arXiv CSopen access

NEXUS-MI: Communication-Aware Federated Personalization for Gateway-Coordinated Motor-Imagery Brain-Computer Interfaces

NEXUS-MI: Communication-Aware Federated Personalization for Gateway-Coordinated Motor-Imagery Brain-Computer Interfaces

arXiv:2609.09786v1 [cs.LG] 9 Sep 2026

Daniel Adu Worae University of Notre Dame [email protected]

Aarthy Nagarajan University of Notre Dame [email protected]

However, the transition from controlled MI-BCI experiments to reliable home use remains limited by decoder stability. EEGbased MI signals are weak, non-stationary, and highly userdependent, with substantial variation across subjects, sessions, and recording conditions [4], [5]. Decoders are affected by subject physiology, electrode placement, attention, fatigue, and session-to-session drift [6]–[8]. A model that performs well during calibration may degrade when used by the same subject in a later session or transferred to another user. In MI-BCI rehabilitation, decoder instability has consequences beyond reduced classification performance: inconsistent predictions can weaken the correspondence between motor imagery and therapeutic feedback, disrupt feedback-driven practice, and increase the need for supervised recalibration. This risk is especially important for users with low or unstable MI-BCI control, for whom further degradation may make the resulting feedback too unreliable to support continued use [6], [9]. Maintaining dependable decoder performance for each user across sessions in deployment-oriented MI-BCI systems is, therefore, a central requirement for sustained rehabilitation use. Existing adaptation strategies address this reliability challenge only partially. Subject-specific training can adapt to individual EEG structure; however, in routine home use, repeatedly collecting large labeled calibration sets is impractical. Training a deep decoder from small calibration sets risks overfitting, while lightweight adaptation may be limited when the representation being adapted is weak or stale. Centralized multi-user training can exploit population-level EEG structure, I. I NTRODUCTION but it requires moving sensitive neurophysiological recordings Motor-imagery brain-computer interfaces (MI-BCIs) offer away from the user environment. Federated learning (FL) offers a promising path toward neurorehabilitation systems that can a natural alternative: multiple users can contribute to improving translate imagined movement into control signals without requir- a shared model while keeping raw EEG data local. Prior MIing overt motor execution [1], [2]. This capability is especially BCI FL studies have shown that collaborative learning can important for longitudinal rehabilitation, where users may improve motor-imagery classification and address cross-user need repeated practice beyond specialized clinical facilities [3]. heterogeneity without centralizing EEG recordings [10]–[12]. These studies establish an important foundation for distributed Daniel Adu Worae is with the Department of Computer Science and Engineering, University of Notre Dame, 307 Fitzpatrick Hall, Notre Dame, MI learning, but they largely evaluate FL as a modeling protocol IN 46556, USA (e-mail: [email protected]). rather than as a gateway-coordinated deployment workflow. Aarthy Nagarajan is with the Department of Computer Science and This distinction matters because home MI-BCI is also Engineering and the Lucy Family Institute for Data and Society, University of an edge coordination problem. A target deployment model Notre Dame, Notre Dame, IN 46556, USA (e-mail: [email protected]). Daniel Adu Worae and Aarthy Nagarajan are corresponding authors. includes a wearable sensor, a subject-side gateway, and an edge

Abstract—Electroencephalography (EEG)-based motor-imagery brain-computer interfaces (MI-BCIs) exhibit variability across subjects and sessions, complicating personalization from limited calibration data. Federated learning can exploit shared representations without centralizing raw EEG. However, existing federated MI studies assume regular synchronization and overlook intermittent gateway availability and coordination. We introduce NEXUSMI, a gateway-coordinated federated personalization framework that formulates synchronization as a coupled learning-andcommunication control problem. During the gateway-coordinated phase, raw EEG and classifier heads remain local while an edge coordinator maintains the shared backbone. We evaluate NEXUS-MI through offline replay using BCI Competition IV Dataset 2a (BCICIV-2a; 9 subjects, 4 classes) and OpenBMI (54 subjects, 2 classes). Session 1 supports backbone learning, and Session 2 provides limited-calibration personalization and held-out testing. An ideal-link reference and six heterogeneouslink policies characterize gateway participation, buffering, staleupdate admission, and backbone-download control. The principal comparison holds delayed-update handling fixed while contrasting non-adaptive and communication-aware synchronization. Paired subject-level comparisons use Holm adjustment, and robustness across five matched realizations is assessed by hierarchical bootstrap. Communication-aware coordination reduced server-toclient backbone traffic by approximately 42% on both datasets, while cohort-level accuracy differences were small and realizationdependent. Cohort averages also concealed subject-level vulnerability, with losses reaching approximately 12 percentage points on BCICIV-2a relative to the ideal-link reference. These findings establish gateway synchronization as an explicit design variable in federated MI personalization and motivate joint evaluation of personalized accuracy, communication cost, update freshness, and subject-level reliability.

coordinator. The gateway performs local inference and local the communication policy affects only the synchronization cost personalization close to the user, while the edge coordinator or also changes the decoding performance achieved through maintains a shared representation across users. In this setting, collaborative personalization. RQ3: Which policies preserve subject-level reliability the gateway-to-edge link is not a passive transport channel. Gateways may be intermittently online, bandwidth may vary under heterogeneous links? We examine whether aggregate across homes, updates may be delayed, and local personal- accuracy hides uneven effects across users by measuring ization may continue from a stale backbone while a gateway worst-subject performance, threshold violations, per-subject waits for a refreshed model. These events influence which users deviations from the ideal-link reference, and sensitivity to link contribute to the shared backbone, how fresh their updates are availability. We evaluate the framework on two public MI datasets when they arrive, which model version each gateway uses locally, and how quickly improved representations reach the with complementary characteristics: BCICIV-2a, a four-class user. Consequently, deployment-oriented MI-BCI evaluation benchmark with nine subjects and two sessions, and OpenBMI, must consider not only whether collaborative learning improves a larger two-class benchmark with 54 subjects and two sessions decoding, but also whether the personalized decoder available [9], [13]. The session structure supports a deployment-oriented to each subject remains reliable when gateway participation, protocol: earlier session data support shared-backbone learning, update delivery, and backbone synchronization are constrained. while later session data emulate continued use with limited Current MI-BCI FL evaluations do not sufficiently expose calibration and held-out testing. Across datasets and personthis deployment dimension. They typically show that federation alization regimes, we compare an ideal-link reference with can support distributed EEG model training, but they do not six heterogeneous-link communication policies. The evaluation treat gateway participation, buffering, stale-update admission, reports on accuracy, client-to-server traffic, server-to-client and backbone download control as first-order factors in per- backbone synchronization, accepted-update staleness, dropped sonalized decoding. As a result, the accuracy-communication or rejected updates, avoided downloads, link-availability sensitradeoff remains under-specified for home rehabilitation, where tivity, and subject-level reliability. The results show that gateway communication policy malimited calibration, session drift, subject-level variability, and heterogeneous connectivity can jointly determine whether per- terially changes the operating point of federated MI-BCI sonalized decoding performance is preserved after deployment. personalization. In the strongest observed setting, P5 increased The gap is therefore specific: MI-BCI FL needs a deployment- mean OpenBMI EIB-PH accuracy by 0.99 percentage points centered evaluation in which communication policy is treated relative to matched P3 while reducing server-to-client backbone as a determinant of personalized decoding performance, not traffic by approximately 41%. However, accuracy effects varied across datasets and personalization regimes, with BCICIV-2a only as a background systems cost. Accordingly, this study examines how gateway synchroniza- exhibiting greater subject-level sensitivity under constrained tion shapes decoding accuracy, communication cost, update communication. Across five matched training and gatewayfreshness, and subject-level reliability in federated MI-BCI availability realizations, the server-to-client traffic reduction personalization under limited calibration and heterogeneous remained approximately 42% on both datasets, whereas the P5– connectivity. Unlike prior MI-BCI FL evaluations that largely P3 accuracy differences were small, varied in direction, and had report classification performance under regular federation, crossed hierarchical-bootstrap 95% confidence intervals that NEXUS-MI is positioned not as a general-purpose FL op- included zero. These findings show that deployment-oriented timizer, but as a deployment-centered framework that makes MI-BCI evaluation should account for synchronization costs, synchronization behavior experimentally testable by relating update freshness, and reliability across users, rather than mean gateway participation, buffering, delayed-update admission, decoding accuracy alone. This paper makes the following contributions: and backbone-download control to the personalized decoder ultimately available to each subject, with outcomes evaluated • Gateway-coordinated MI-BCI personalization frameagainst an ideal-link reference. work. We introduce an edge-enabled MI-BCI framework The study is organized around three research questions. that separates subject-local classifier-head adaptation RQ1: How well does gateway-coordinated personalization from cohort-level backbone learning, enabling limitedperform under ideal synchronization? We first establish calibration personalization while retaining raw EEG and an ideal-link reference for SB-PH and EIB-PH. This consubject-specific heads at the gateway during the gatewaydition establishes a reference learning and synchronization coordinated phase. trajectory under unconstrained gateway-to-edge communication • Communication-policy formulation for MI-BCI FL. and provides the baseline needed to interpret communicationWe formulate gateway synchronization as a configurable constrained operation. policy space spanning gateway participation, buffering, RQ2: How do communication policies change the stale-update admission, and version-aware backbone downaccuracy-communication tradeoff? We evaluate gateway-toload control. This formulation makes the coordination edge policies that vary participation, buffering, stale-update admechanisms governing shared-backbone exchange explicit mission, communication-aware scheduling, and version-aware design variables rather than implicit implementation backbone download control. This analysis determines whether assumptions.

Controlled deployment-oriented evaluation. We eval- alignment further reinforces the importance of distribution uate NEXUS-MI on BCICIV-2a and OpenBMI using alignment in cross-subject and cross-session EEG transfer [21]. a session-based protocol that contrasts an ideal-link These methods motivate the use of reusable representations reference with six policies under heterogeneous gateway and limited target-subject adaptation. Beyond alignment-based availability and matched personalization regimes. This transfer, EEG representation pretraining can further reduce calicontrolled comparison quantifies how synchronization bration burden by providing a stronger backbone before subjectpolicy affects personalized decoding, communication specific adaptation. Prior work has shown that pretrained EEG traffic, update freshness, delayed-update rejection, and representations can improve learning when labeled target data avoided backbone downloads. are limited and can transfer across subjects, datasets, and downstream tasks [22], [23]. More recent large-scale pretrained • Subject-level reliability analysis. We show how aggregate accuracy can obscure uneven subject-level effects under models, including LaBraM, EEGPT, and CBraMod, extend this constrained communication by measuring worst-subject principle through self-supervised objectives designed to learn performance, threshold violations, per-subject deviation transferable representations from heterogeneous EEG data [24]– from the ideal-link reference, and link-availability sensitiv- [26]. MIRepNet complements this general-purpose direction ity. This exposes whether communication-efficient policies with an MI-specific pretrained pipeline designed for rapid preserve reliable decoding across users. adaptation under small calibration budgets [27]. Collectively, The rest of the paper is organized as follows. Section II these developments support evaluating whether a backbone reviews related work. Section III presents the system model initialized from pooled Session-1 EEG can facilitate later headand problem formulation. Section IV describes the gateway- only personalization with limited target-session data. From a neural-engineering perspective, sustained MI-BCI coordinated federated personalization framework. Section V use also requires reliable decoding across sessions with limited presents the experimental evaluation. Section VI discusses recalibration, particularly for users with low or variable deployment implications. Section VII concludes the paper. control [6], [9], [28]. Recent healthy-to-stroke transfer results further indicate that pretrained MI representations can support II. R ELATED W ORK adaptation when clinical data are limited [29]. Wearable and A. MI-BCI Decoding, Calibration, and Personalization edge-enabled BCI systems further motivate local, resourceMotor-imagery EEG decoding has a long history of methods aware operation beyond controlled laboratory settings [3], that seek discriminative sensorimotor-rhythm structure from [30], [31]. Existing work, however, largely studies calibration noisy, low-SNR signals. Classical pipelines based on common reduction through transfer, alignment, representation initialspatial patterns and filter-bank common spatial patterns re- ization, or target-subject adaptation rather than how a shared main influential because they combine spatial filtering with EEG representation is maintained through intermittent gatewayfrequency-specific MI information [14]. Deep EEG models to-edge synchronization and subsequently reused for local extend this tradition by learning temporal, spectral, and spatial personalization. representations directly from data. Shallow and deep convolutional EEG decoders, compact architectures such as EEGNet, B. Federated and Personalized Learning for EEG and BCI and MI-focused models such as FBCNet have established strong Federated learning changes the collaboration model by neural backbones for MI classification under limited-sample allowing distributed users or sites to improve a model without conditions [15]–[17]. More recent work has continued this centralizing raw EEG data. In EEG and BCI, federated transfer progression by integrating established spatial priors with end-to- learning has shown that cross-subject information can be end representation learning; CSP-Net, for example, incorporates exploited in a distributed setting [11]. Recent MI-BCI federated common-spatial-pattern structure into neural models for both learning further addresses non-IID client data through clientsubject-dependent and subject-independent MI decoding [18]. specific normalization and robust local optimization [10]. Other These works improve the representational substrate for MI federated BCI work studies collaboration across heterogeneous decoding, but they do not address how such representations EEG datasets and acquisition devices, showing that federation should be maintained across users when the deployment system can improve performance when data sources differ in format, is distributed across gateways and an edge coordinator. scale, and collection conditions [12]. More recent approaches A second line of work reduces calibration burden through extend this direction through structure-guided personalization transfer learning, domain alignment, and subject-adaptive for heterogeneous MI clients and cross-subject federation decoding. Transfer learning is widely used in EEG-based designed to improve generalization to unseen EEG users [32], BCI because EEG distributions vary across subjects, sessions, [33]. Together, these works establish FL as a valuable learning devices, and tasks [5]. Alignment-based approaches make paradigm for EEG and MI-BCI. source and target EEG data more comparable before feature The limitation is that most EEG/BCI FL studies frame extraction or model adaptation, while MI-specific transfer the problem primarily as distributed model training. The pipelines combine alignment, spatial filtering, feature learning, communication process is usually represented as a standard and classifier adaptation to reduce the amount of new-user cal- federation protocol rather than as a deployment variable that ibration required [19], [20]. Recent reassessment of Euclidean can alter personalized decoding. Questions such as which •

gateways participate, whether delayed updates should be while also increasing the risk of stale contributions. Therefore, buffered, when stale updates should be admitted, and whether communication mechanisms must be evaluated not only as backbone downloads should be avoided are largely outside the FL systems optimizations, but as determinants of MI-BCI evaluation scope. This leaves an important gap for home MI- personalization quality. BCI, where gateway availability and synchronization behavior can determine which model version a subject uses and how D. Positioning of This Work This paper connects the above threads in a deploymentfresh the shared representation is during personalization. Personalized FL provides a useful modeling foundation for oriented MI-BCI setting. MI decoding and transfer-learning this gap. Methods based on personalization layers, meta-learned studies provide strong backbones and motivate limitedinitialization, and shared representations with local heads calibration adaptation. EEG/BCI FL studies show that colshow that client heterogeneity can be handled by separating laboration can improve decoding while keeping raw EEG local. global and local model components [34]–[36]. This idea aligns Personalized FL motivates the shared-backbone and subjectnaturally with MI-BCI: a shared backbone can encode reusable specific-head structure. Communication-aware FL systems proEEG structure, while a subject-specific head can adapt the vide mechanisms for participation control, buffering, staleness decision boundary from limited calibration data. Our work handling, and edge coordination. The distinction is that this paper evaluates these mechanisms adopts this principle, but evaluates it in a gateway-coordinated setting where the shared backbone is not assumed to be through the requirements of gateway-coordinated MI-BCI synchronized ideally. The central issue is not only how to personalization. Unlike model-centered MI-BCI FL work, it personalize, but whether personalization remains reliable when treats synchronization behavior as part of the learning system the shared representation is refreshed through heterogeneous rather than a fixed background assumption. Unlike generic communication-efficient FL work, it evaluates communication and policy-controlled communication. policy using MI-BCI outcomes: session-based personalization C. Communication-Aware, Asynchronous, and Edge Federated accuracy, client-to-server update cost, server-to-client backbone Learning synchronization, update staleness, dropped updates, avoided The FL systems literature has developed a rich set of downloads, link-availability sensitivity, and subject-level reliamechanisms for reducing communication cost and handling bility. This positioning makes the work a deployment-centered heterogeneous clients. FedAvg established iterative model aver- study of federated MI-BCI personalization, in which learning aging as a practical foundation for decentralized model learning performance and communication behavior are evaluated jointly. In NEXUS-MI, data locality applies to the gateway[37]. FedProx addresses statistical and systems heterogeneity by modifying the local optimization objective [38]. Adaptive coordinated personalization and synchronization phase; EIBFL controls the tradeoff between local computation and global PH uses pooled Session-1 data for predeployment backbone aggregation under edge resource constraints [39]. Client- initialization, while formal privacy mechanisms such as secure selection systems such as Oort prioritize participants based on aggregation, differential privacy, and encrypted computation utility and system capability to improve time-to-accuracy [40]. are not evaluated. Asynchronous and buffered approaches, including FedAsyncIII. S YSTEM M ODEL AND P ROBLEM F ORMULATION style optimization and FedBuff, reduce the need to wait for fully synchronized rounds and provide mechanisms for A. Gateway-Coordinated MI-BCI Setting handling delayed updates [41], [42]. More recent systems We consider an edge-enabled motor-imagery brain-computer combine compute-aware semi-asynchronous scheduling with interface (MI-BCI) system with a set of subjects S = on-demand model broadcasting to reduce straggler delays, {1, . . . , N }, where N is the number of subjects. Each subject update staleness, and communication under heterogeneous or s ∈ S is served by a local gateway connected to a wearable personalized deployments [43], [44]. Hierarchical FL further EEG acquisition device. The gateway performs local inference introduces intermediate edge aggregation to reduce cloud for latency-sensitive decoding, stores the subject’s EEG trials backhaul, latency, and end-device energy cost [45], [46]. and calibration labels, and carries out subject-specific adapThese systems contributions are closely related to the tation. During the gateway-coordinated personalization and mechanics of gateway-based MI-BCI, but their objectives are synchronization phase, raw EEG trials, calibration labels, and different. They are usually evaluated through convergence subject-specific classifier heads remain at the gateway. Each behavior, time-to-accuracy, communication rounds, scalability, gateway uploads a locally trained backbone parameter delta or aggregate model performance on general machine-learning together with the metadata required for version-aware admission workloads. MI-BCI deployment imposes a narrower and and reconstruction, while the coordinator may return a refreshed more demanding interpretation of communication efficiency. backbone state. The edge coordinator maintains the cohortA policy that reduces traffic can still be unacceptable if it level shared backbone and coordinates model synchronization reduces the reliability of feedback for subjects with weak across gateways. links. A policy that maximizes freshness can be impractical Because this study uses a one-subject/one-gateway abstracif it requires frequent server-to-client backbone transfers. A tion, subject and gateway indices are in one-to-one corresponpolicy that admits delayed updates can improve participation dence. We use s for subject data and decoder quantities, and i

only when emphasizing the communication state. Thus, subject(1) (2) indexed quantities such as Ds , Ds , and ϕs describe the user’s data and personalized classifier head, whereas gatewayindexed quantities such as ari , vir , and communication buffers describe synchronization state. The gateway is the operational unit of the deployment. For each subject s, all subject-specific data and decoder components remain at the subject-side gateway. The edge coordinator does not receive raw EEG trials, calibration labels, or personalized classifier heads. Instead, it receives only backbone-related information produced by local training. This design allows each gateway to adapt to its subject while still allowing the shared representation to benefit from multi-subject learning. The resulting system is neither purely local nor fully centralized: gateways retain subject-specific data and personalization, while the edge coordinator maintains the shared representation used across the cohort. In the present evaluation, this architecture is instantiated as a deployment model over public session-based MI-EEG datasets rather than as a real-time wearable implementation. Each subject is treated as a gateway-associated user, and recorded sessions are replayed to study how synchronization policy affects later-session personalization. B. Session-Based Personalization Protocol The system follows a session-based protocol that reflects repeated MI-BCI use. For each subject s, the available data are divided into an earlier session and a later session. The earlier session provides data for learning a shared representation and an initial subject-specific decoder state. The later session emulates continued use, where only a limited number of labeled calibration trials are available before evaluating the personalized decoder on held-out trials. Let D(1) s denote the earlier-session data for subject s, and let D(2) s denote the later-session data. From D(2) s, the gateway forms a limited calibration set Cs, k with calibration budget k, and a disjoint held-out evaluation set T s, k. The calibration set is used for subject-local personalization at the gateway. No held-out trials from T s, k are used during gateway-side personalization. The held-out set is used only to measure latersession decoding after personalization. This protocol separates two learning roles. The earlier session supports cohort-level representation learning across subjects. The later session tests whether that representation remains useful when each gateway performs limited-calibration personalization for its own subject. The protocol, therefore, evaluates continued-use personalization rather than only training-time classification performance. C. Backbone–Head Model Decomposition Each subject-specific decoder is decomposed into a shared EEG feature backbone and a personalized classifier head. Let gθ denote the backbone with parameters θ, and let hϕs denote the classifier head for subject s, with parameters ϕs . The local decoder at gateway s is
fs (x) = hϕs gθs (x) , (1)

where θs denotes the backbone state currently available at the gateway serving subject s. The backbone captures EEG representations that can benefit from multi-subject learning, whereas the classifier head represents the subject-specific decision boundary. In the evaluated EEGNet configuration, gθ comprises the two feature-extraction blocks preceding the terminal classifier. The first block applies eight temporal filters with a 1 × 125 kernel, batch normalization, and a depthwise spatial convolution with depth multiplier two spanning all EEG channels, followed by batch normalization, exponential linear activation, 1 × 4 average pooling, and dropout. The second block applies a 1 × 22 depthwise temporal convolution and a 1 × 1 pointwise convolution producing 16 feature maps, followed by batch normalization, exponential linear activation, 1 × 8 average pooling, and dropout. For the 1,000-sample input windows used in both datasets, hϕs consists of the terminal convolution that maps the 16 backbone feature maps to the dataset-specific classes using a 1 × 31 kernel, followed by log-softmax normalization. This decomposition yields 2,040 trainable backbone parameters and 1,988 trainable head parameters for BCICIV-2a, and 2,008 backbone parameters and 994 head parameters for OpenBMI, corresponding to 4,028 and 3,002 trainable parameters in total, respectively. The edge coordinator maintains the shared backbone state, while each gateway maintains a local backbone copy and a subject-specific head. At each Session-1 local update, the gateway combines its currently available backbone with the common initial head state and jointly optimizes both components using the subject’s Session-1 data. Joint optimization allows the subject-specific classification objective to shape the resulting backbone update. After local training, only the backbone-related update and its version metadata are eligible for synchronization with the edge coordinator; the trained classifier head remains at the gateway and replaces the previously retained head. Successive Session-1 local updates reuse the common head initialization. This limits carryover of head–backbone co-adaptation across communication rounds, so each backbone update is learned from the gateway’s current backbone state under a consistent classifier initialization rather than from a classifier shaped by earlier backbone states and participation history. The most recently trained head is retained locally to initialize Session-2 personalization. During Session-2 personalization, each gateway combines the final collaborative backbone with the subject’s most recently retained Session-1 head and adapts that head using Cs,k . The backbone’s trainable parameters are held fixed during this stage, and only the classifier-head parameters are optimized. Personalization therefore begins from a subject-specific head learned during Session 1 rather than from a randomly initialized classifier. This decomposition is central to the system model because it separates cohort-level representation learning from subjectlocal adaptation. A gateway communication policy can alter the backbone version available locally, the freshness of updates reaching the coordinator, and the frequency with which

refreshed backbone states return to the gateway. These effects can influence subsequent personalization even when the local calibration budget and the underlying EEG data remain fixed. D. Gateway-to-Edge Communication Model

calibration budget, each communication policy π induces the operating point  z(π) = {As (π)}s∈S , C ↑ (π), C ↓ (π), Rrej (π),  (2) Lacc (π), Dbuf (π), {∆s (π)}s∈S ,

Communication occurs over synchronization rounds. At where As (π) is the held-out later-session accuracy of subject s; ↑ ↓ round r, the edge coordinator maintains a global backbone C (π) and C (π) are the total client-to-server model-update r r version V with parameters θ . Gateway i maintains a local traffic and server-to-client backbone traffic; Rrej (π) is the backbone version vir with parameters θir . The difference V r −vir fraction of transmitted uploads rejected by the coordinator represents the gateway’s version lag relative to the coordinator. because their base checkpoint is unavailable or their backboneGateway availability is heterogeneous and time-varying. Let version lag exceeds the stale-update threshold; Lacc (π) is the ari ∈ 0, 1 denote the gateway-to-edge link state for gateway i at event-weighted mean backbone-version lag of updates that pass round r. If ari = 1, the gateway is available for synchronization coordinator admission and enter aggregation; and Dbuf (π) with the edge coordinator. If ari = 0, the gateway is unavailable is the event-weighted mean communication-round delay of for edge communication in that round. An unavailable gateway buffered uploads that pass the checkpoint-availability check may continue local operation using its cached model state, but and reach stale-update admission. Subject-level change relative it cannot upload an update or download a refreshed backbone to the ideal-link reference is defined as until communication becomes available. ∆s (π) = As (π) − Aideal . (3) s The communication loop has two directions. In the clientThis formulation treats gateway coordination as part of the to-server direction, a gateway uploads a backbone update expressed relative to the backbone version from which local federated MI-BCI learning system rather than as an independent training began, together with metadata identifying that base communication overhead. The policies are compared as alterversion. These uploads determine which subjects contribute native operating points rather than optimized through a single to the next shared backbone. In the server-to-client direction, scalar objective, because communication savings, coordinator the coordinator sends a refreshed backbone state to a gate- rejection, accepted-update staleness, and buffered-upload delay way. These downloads determine which backbone version is may have different accuracy effects across datasets, learning available for subsequent local training and personalization. regimes, and subjects. Subject-level reliability is assessed using Subject-specific classifier heads are not exchanged in this loop; both absolute personalized accuracy and change relative to the ideal-link reference, since aggregate performance can conceal they remain local to the gateways. Intermittent communication can desynchronize the backbone substantial degradation or persistently low decoding accuracy versions held by the coordinator and the gateways. An update for individual users. computed at a gateway may reach the coordinator after the IV. G ATEWAY-C OORDINATED F EDERATED global backbone has advanced, and a gateway may continue P ERSONALIZATION local operation using a cached backbone until a newer shared NEXUS-MI operationalizes gateway-coordinated MI-BCI state is received. These conditions define the version-lag and personalization through two design choices: a backbone–head update-freshness variables used by the coordination policies learning structure and a communication policy that governs in Section IV; the specific rules for buffering, stale-update synchronization between gateways and the edge coordinator. admission, and backbone downloads are introduced there. The learning structure determines how cohort-level representaWe denote a gateway coordination policy by π. At each tion learning and subject-local adaptation are separated. The round, π acts on the system variables defined above, including communication policy determines how backbone updates are gateway availability, local and global backbone versions, delivered, admitted, and redistributed when gateway availability pending updates, and gateway participation. The specific is heterogeneous. Section IV-A defines the federated personpolicy mechanisms are defined in Section IV; this section alization regimes; Section IV-B defines the synchronization establishes the general variables and constraints on which conditions; Section IV-C summarizes the policy taxonomy; those mechanisms operate. Section IV-D specifies the six gateway coordination policies; Section IV-E details communication-aware gateway scheduling; Section IV-F describes buffering, checkpoint retention, E. Problem Statement and stale-update admission; and Section IV-G defines staleConsider a set of subject-side gateways with local MI- aware backbone downloads. The framework builds on estabEEG sessions, limited later-session calibration data, and an lished federated-learning mechanisms for gateway participation, edge coordinator that maintains a shared EEG backbone buffering, stale-update handling, and version-aware model under heterogeneous gateway-to-edge connectivity. We study delivery. Its contribution is to organize these mechanisms how gateway-coordination policy affects personalized MI- as a coordinated policy space for MI-BCI personalization BCI performance. For a fixed dataset, learning regime, and and to evaluate how their combinations shape later-session

decoding accuracy, communication cost, update freshness, and subject-level reliability under heterogeneous gateway-to-edge connectivity, rather than treating synchronization as a fixed communication assumption. A. Federated Personalization Regimes

initial head state. This models continued decoder use, in which later-session adaptation begins from an existing subject-specific state produced during Session 1 training whenever such a state is available.

B. Synchronization Conditions We evaluate two collaborative personalization regimes based The gateway-coordinated personalization framework is evalon the decoder decomposition in Eq. 1. In both regimes, the uated under an ideal-link reference and under heterogeneous backbone provides the shared EEG representation, while the gateway-to-edge connectivity. The ideal-link reference uses classifier head remains subject-specific and is personalized at the same collaborative personalization pipeline but removes the gateway during later-session adaptation. communication-induced constraints: selected gateways are Shared Backbone with Personalized Head (SB-PH). SB- available, synchronize with the edge coordinator when required, PH learns a shared backbone through collaborative training train locally, and upload current-round updates without missed across subject gateways. After collaborative backbone learning, uploads, delayed delivery, stale-update rejection, or baseeach gateway personalizes only its classifier head. This regime checkpoint unavailability caused by link loss. evaluates whether a cohort-learned representation can support Under heterogeneous connectivity, each gateway i is assigned limited-calibration subject adaptation while retaining a local once to an availability group γ ∈ {high, moderate, low} with i decision boundary. online probability pγi . At each communication round r, the Embedding-Initialized Backbone with Personalized Head gateway’s online/offline state is sampled independently across (EIB-PH). EIB-PH evaluates whether gateway-coordinated gateways and rounds as personalization benefits from a stronger initial EEG representation. In the present evaluation, the backbone is first initialized ari ∼ Bernoulli(pγi ), (4) in a pre-deployment representation-learning step using pooled r r Session 1 data from the training subjects. The learned backbone where ai = 1 denotes that gateway i is online and ai = 0 weights are then used to initialize the subsequent gateway- denotes that it is offline. For matched policy comparisons, the coordinated training, communication-policy evaluation, and same generated availability trace is used across policies, so later-session head-only personalization procedure. Thus, EIB- each policy is evaluated under the same realized sequence of PH changes the starting representation, but not the later-session online and offline gateway states. The sampled state affects calibration budget, personalization rule, communication policy, gateway-edge synchronization only. An offline gateway retains its cached model state and cannot upload an update or download or evaluation protocol. After this initialization step, EIB-PH follows the same a refreshed backbone until a later online round. When the active gateway-local structure as SB-PH. Raw EEG trials, calibration policy enables offline buffering, the gateway may also retain a labels, and subject-specific classifier heads remain at the subject- locally produced pending update for later synchronization. The main heterogeneous-link policy study uses the default side gateways during the gateway-coordinated phase; the edge availability setting: high-, moderate-, and low-availability coordinator exchanges only backbone-related updates, version gateways have online probabilities 0.95, 0.70, and 0.40, metadata, and policy-level synchronization state. The pooled respectively, with target group fractions 0.34, 0.33, and 0.33. initialization, therefore, defines the starting representation Because each dataset contains a finite number of subject for EIB-PH, while the subsequent experiments evaluate how gateways, these fractions define the intended group proportions, gateway synchronization policies affect personalization from while the realized group counts are determined by the available that common initialization. subject count. These operating points are intended to emulate This regime separates two methodological questions. SBreliable home or institutional Wi-Fi, intermittent consumer PH measures how well a shared backbone can be learned connectivity, and weak or mobile backhaul. This abstraction directly through the federated personalization pipeline. EIB-PH follows prior cross-device FL work on availability, participant measures whether initializing that backbone with a stronger selection, and delayed aggregation [40], [42], [47]. cohort-level representation improves subsequent personalization To avoid tying the conclusions to a single availability and changes sensitivity to gateway synchronization. Reporting setting, the sensitivity study repeats the same high/moderate/low SB-PH and EIB-PH separately prevents conflating the effect of grouping under three severity levels. The mild setting uses representation initialization with that of gateway coordination. online probabilities 0.98/0.85/0.60, the default setting uses For both regimes, later-session personalization starts from the 0.95/0.70/0.40, and the severe setting uses 0.90/0.50/0.20 for final collaborative backbone. During Session 2 personalization, the high-, moderate-, and low-availability groups, respectively. the gateway keeps this backbone fixed and updates only the subject-specific classifier head using the limited calibration set Cs,k . For each subject, the classifier head is initialized from the C. Policy Taxonomy latest locally stored Session 1 head produced by that subject All six heterogeneous-link settings are gateway coordination during collaborative training. If no Session 1 head state is policies. They share the same link model but differ in four available for a subject, personalization falls back to the common synchronization dimensions: scheduling, offline update rule,

2) GATEWAY LAYER (Subject Side) Local Inference, Personalization, and Communication

A. System architecture

EEG preprocessing

Cached shared backbone

3) EDGE COORDINATOR LAYER

Cohort-Level Orchestration and Aggregation

Local inference

Policy-controlled gateway scheduling

2

Checkpoint- and staleness-aware upload admission

3

Accepted-update reconstruction + aggregation

4

Backbone checkpoint retention + version tracking

5

Policy-controlled backbone distribution

Backbone delta + base-version metadata

1) WEARABLE LAYER (Subject Side)

EEG Wearable

1

Short-range local link (BLE/Wi-Fi) Session-1 local update

Raw EEG stream

Limited-calibration head adaptation

Personalized classifier head

Buffered while offline, when enabled

MI-EEG Acquisition

Subject 2 Local EEG + calibration data store

Subject N

Pending-update buffer

Gateway link state

Shared backbone refresh (always or versionlag-aware)

Raw EEG, labels, and personalized heads remain local

B. Experimental instantiation Session-Based Protocol

Communication Policy

1

1

Session 1: collaborative backbone learning + local head-state retention

2

Session 2: limited-calibration personalization + held-out evaluation

Replay of public datasets

Legend:

Gateway scheduling

2

Offline-update handling

3

Stale-update admission

4

Local flow (on-device / short-range link)

Regime Summary

Backbone download rule

SB-PH:

Common initialization

Federated backbone learning

Local head personalization

EIB-PH:

Pooled Session1 preinitialization

Federated backbone learning

Local head personalization

Network synchronization (internet)

Remains gateway-local (data locality)

Fig. 1. System architecture and experimental instantiation of NEXUS-MI. (A) The system follows a three-tier wearable–gateway–edge design for communicationaware federated MI-BCI personalization. The wearable acquires MI-EEG signals and streams them to a subject-side gateway, where EEG preprocessing, local inference, Session-1 local updates, limited-calibration head adaptation, local data storage, link monitoring, and update buffering are performed. Raw EEG, calibration labels, and personalized classifier heads remain local to the gateway. The gateway communicates backbone deltas and associated base-version metadata to the edge coordinator, which performs policy-controlled gateway scheduling, checkpoint- and staleness-aware upload admission, accepted-update reconstruction and aggregation, backbone checkpoint retention and version tracking, and policy-controlled backbone distribution according to the active download rule. (B) The architecture in (A) represents the intended deployment workflow, whereas the experiments instantiate this workflow through replay of public session-based MI-EEG datasets rather than a live online deployment. Session 1 supports collaborative backbone learning, while Session 2 uses limited calibration for local head personalization followed by held-out evaluation. The communication policy is defined by four synchronization dimensions: gateway scheduling, offline-update handling, stale-update admission, and backbone download control; the detailed policy configurations are provided in Table I. SB-PH follows common initialization, federated backbone learning, and local head personalization, whereas EIB-PH uses pooled Session-1 preinitialization followed by the same federated backbone-learning and local head-personalization pipeline.

stale update rule, and download rule. Table I gives the compact gateway produces an update but cannot immediately upload definitions used throughout the paper. it. The update may be discarded, stored in a bounded FIFO The scheduling dimension determines which gateways buffer, or used to replace the gateway’s previous pending receive immediate upload opportunities. Under non-adaptive update under the latest-update rule. The stale update rule all-gateway scheduling, every gateway is scheduled in each determines whether a delayed update is admitted after it communication round before the realized link state is applied. later reaches the coordinator. The update may be rejected Under communication-aware gateway scheduling, the coordina- because no delayed update was retained, admitted if its base tor first observes the realized online set and assigns immediate checkpoint is still available, or rejected if its version lag upload opportunities only to gateways that can communicate exceeds the stale threshold τs . The download rule determines whether an online gateway receives a refreshed backbone in that round. The offline update rule determines what happens when a from the coordinator. Always-download synchronization sends

TABLE I G ATEWAY COORDINATION POLICIES UNDER HETEROGENEOUS GATEWAY- TO - EDGE LINKS .

Policy

Scheduling

Offline update rule

Stale update rule

Download rule

P1 P2 P3 P4 P5 P6

NA-all NA-all NA-all NA-all CA-priority CA-priority

Discard missed update FIFO buffer FIFO buffer Latest pending update FIFO buffer Latest pending update

No delayed update Accept if base retained Drop if ℓru > τs Drop if ℓru > τs Drop if ℓru > τs Drop if ℓru > τs

Download when selected Download when selected Download when selected Download when selected Refresh if V r − vir > τd Refresh if V r − vir > τd

Notes: NA-all denotes non-adaptive all-gateway scheduling, in which gateways are selected before online/offline availability is realized. CA-priority denotes communication-aware scheduling over currently online gateways using the priority rule in Eq. 6. FIFO buffers are bounded with Bmax = 3 in the experiments. The latest pending update retains only the most recent delayed update for a gateway. The stale-update threshold is τs = 2 and is applied to the delayed-update lag ℓru defined in Eq. 9. The download threshold is τd = 1 and is applied to the gateway’s local backbone lag. P3 and P5 share the same FIFO offline-update rule and stale-update rule; they differ only in the scheduling and download-rule columns.

the newer backbone whenever one is available, while staleaware synchronization sends it only when the gateway’s cached backbone is sufficiently behind the coordinator. D. Gateway Coordination Policies

Only gateways in Or are eligible for immediate upload. Let K denote the per-round gateway-selection budget; in the experimental protocol, K = 6 for BCICIV-2a and K = 40 for OpenBMI, as summarized in Table II. If |Or | ≤ K, all online gateways are selected. If |Or | > K, the coordinator ranks online gateways by synchronization need and selects the highest-priority gateways. The priority ordering is a fixed scheduling heuristic rather than the solution of an explicit optimization problem; its intuition is to favor gateways that have not uploaded successfully recently, carry buffered updates at greater risk of becoming stale, or hold a more outdated local backbone. For each online gateway i, the coordinator computes the priority vector  succ  ∆i (r)  Tisucc (r)     Bi (r)   . qi (r) =  buf (6)   Sibuf(r)   Di (r)  ∆loc i (r)

The policies form two families. P1–P4 provide the nonadaptive reference family under heterogeneous gateway availability. In these policies, all gateways are scheduled before the online/offline state is applied; only gateways that are online in the realized link state can synchronize in that round. Differences within this family therefore isolate how missed updates, buffering, and stale-update admission affect collaborative backbone learning when scheduling itself does not adapt to availability. P5–P6 provide the communication-aware family. These policies first observe the currently online gateway set and assign immediate upload opportunities only to gateways that can communicate in the current round. They also use staleaware backbone downloading, so a gateway refreshes its cached backbone only when its local version lag exceeds the download threshold. The difference between P5 and P6 is the offlineupdate rule: P5 retains delayed updates in a bounded FIFO Here, ∆succ (r) is the coordinator-version lag since gateway i buffer, whereas P6 retains only the latest pending update from i’s last successful upload, Tisucc (r) is the number of rounds each offline gateway. since that upload, Bi (r) indicates whether the gateway has a This policy organization makes P3 and P5 the principal nonempty pending-update buffer, Sibuf (r) and Dibuf (r) are the controlled comparison. Both policies use FIFO buffering and maximum staleness and delay among buffered payloads, and stale-update rejection, so delayed-update handling is held fixed. ∆loc i (r) is the lag between the current coordinator backbone They differ in synchronization behavior: P3 uses non-adaptive and the gateway’s cached local backbone. all-gateway scheduling with always-download synchronizaThe coordinator sorts online gateways in descending lexicotion, whereas P5 combines communication-aware gateway graphic order of qi (r). At the start of training, gateways have scheduling with stale-aware backbone downloading. The P3/P5 no prior successful upload, so the last successful upload round comparison, therefore, isolates the effect of communicationand version are initialized to zero, and the corresponding lag aware synchronization that adapts gateway selection and terms are measured from the beginning of the collaborative backbone refresh to realized availability and version lag while process. For a gateway with no pending update, Bi (r) = 0, holding the buffering and stale-update admission rules constant. Sibuf (r) = −1, and Dibuf (r) = 0. Ties are resolved using a seeded round-specific random key, followed by the subject E. Communication-Aware Gateway Scheduling P5 and P6 use an online-first scheduling rule. At the identifier, so matched policy comparisons remain reproducible. beginning of communication round r, the heterogeneous link Because the ordering is lexicographic, earlier components intentionally dominate later components: version lag since model determines the online and offline gateway sets: the last successful upload is prioritized before time since the Or = {i : ari = 1}, Fr = {i : ari = 0}. (5) last successful upload, pending-buffer status, buffered-update

staleness, buffered-update delay, and local-backbone lag. The number of selected gateways is

finite. After the coordinator advances to version V r , it retains checkpoint versions

mr = min{K, |Or |},

v ≥ V r − τs − ρ.

(7)

and the selected set contains the top mr online gateways under the priority ordering. The scheduling rule defines which gateways perform an immediate upload in a communication round. Under nonadaptive all-gateway policies, every gateway is scheduled before availability is applied. A scheduled online gateway may refresh its cached backbone according to the active download rule, train locally, and upload a fresh update. A scheduled offline gateway trains from its cached backbone, and the resulting update is discarded or retained according to the active offlineupdate rule. Under communication-aware policies, only selected online gateways receive immediate upload opportunities. A selected online gateway first delivers any pending buffered updates and then trains and uploads a fresh update from its current cached or refreshed backbone. Each buffered payload and the subsequently generated fresh payload are transmitted as separate update instances and are evaluated independently by the coordinator. Consequently, a gateway may contribute multiple accepted update instances in the same communication round. Offline gateways are not selected for immediate upload; under P5 and P6, they may continue local training from their cached backbones and retain the resulting updates according to the active buffering rule. Online gateways that are available but not selected within the gateway-selection budget remain idle and do not train, upload, or download in that communication round. The local training objective, update reconstruction, aggregation rule, buffering behavior, stale-update admission, and backbone-download rule remain governed by the active communication policy. F. Buffering, Checkpoints, and Stale-Update Admission

(8)

where τs denotes the stale-update admission threshold and ρ denotes the checkpoint-retention margin. The margin controls reconstruction availability, while the stale-update threshold controls admission. This separation allows delayed updates to be reconstructed from their base checkpoints before the coordinator determines whether they are admissible under the staleness rule. We set ρ = 5 for all experiments and keep it fixed across datasets, subjects, and stale-drop policies. This value provides a reconstruction margin larger than the staleadmission threshold, allowing recently delayed payloads to be reconstructed from their base checkpoints while keeping the retained checkpoint history bounded. With τs = 2, the retained history contains at most eight checkpoint versions. For accept-if-base admission, the checkpoint history is not pruned by the stale-drop horizon, since delayed updates are screened by base-checkpoint availability rather than by staleness. When a delayed update later reaches the coordinator, the coordinator first checks whether the corresponding base checkpoint is still retained. If the base checkpoint is unavailable, the update is rejected as checkpoint-missing. Such drops occur when a delayed upload refers to a base version that has aged out of the retained checkpoint history. If the base checkpoint is available, the update’s staleness is computed as ℓru = V r − vu .

(9)

Here, V r is the coordinator backbone version at round r, and vu is the backbone version from which the delayed update u was trained. Accept-if-base admission admits delayed updates whose base checkpoint is available. Stale-drop admission further rejects delayed updates satisfying

ℓru > τs . (10) When a gateway cannot deliver an update in the current communication round, the active offline update rule determines Stale-drop admission is instantiated with a fixed threshold whether the locally produced update is retained for later τs = 2 for all subjects, allowing modest delayed delivery while synchronization. With no buffering, the missed update is rejecting updates trained from backbone versions that are too far discarded. With FIFO buffering, pending updates are stored up behind the coordinator. For an accepted update u, let vu denote to a bounded capacity Bmax , and the oldest pending update is the coordinator backbone version from which the gateway dropped when the buffer overflows. With latest-update buffering, trained, and let ∆θu denote the transmitted backbone delta. only the most recent pending update is retained for that gateway, The corresponding candidate backbone state is reconstructed replacing any older pending update. Bounded buffering is as instantiated with capacity Bmax = 3, allowing each gateway to θ̃ur = θvu + ∆θu . (11) recover from short offline periods while preventing unbounded queues of delayed and increasingly stale updates. This reconstruction makes each delayed update interpretable A delayed update is interpretable only with respect to the relative to the retained base checkpoint from which it was backbone checkpoint from which it was trained. In the gateway- produced. to-coordinator direction, each communicated learning payload Accepted candidate backbone states are aggregated at the consists of a backbone delta and its associated metadata, update-instance level using a uniform averaging rule that is held including the coordinator version from which the local training fixed across all policies. Because buffered and fresh payloads step began. The coordinator, therefore, maintains a versioned are evaluated independently, a gateway that contributes multiple checkpoint history so that delayed backbone deltas can be accepted payloads in a communication round contributes reconstructed as candidate backbone states before aggregation. multiple candidate backbone states to that round’s aggregation. For policies that apply stale-drop admission, this history is Gateway scheduling, buffering, and coordinator admission

therefore determine the composition of the accepted-update set. The shared backbone is updated as 1 X r θr+1 = θ̃u , |Ar | > 0. (12) |Ar |

cost. In this setting, gateway availability, delayed update delivery, stale-update admission, and backbone synchronization determine which subject updates shape the shared representation and which backbone version is available for subsequent u∈Ar personalization. The evaluation, therefore, reports decoding Here, Ar denotes the set of accepted update instances at round accuracy together with communication volume, update rejecr, and each accepted buffered or fresh payload contributes tion, accepted-update staleness, backbone download behavior, one reconstructed candidate backbone state θ̃ur . As defined in link-availability sensitivity, and subject-level reliability. The analysis proceeds from reference behavior to constrained Eq. (11), each candidate state is reconstructed from the retained deployment behavior. We first establish the ideal-link perforcheckpoint corresponding to the base version identified in the mance of the two personalization regimes, SB-PH and EIB-PH, payload metadata. If Ar is empty, no aggregation is performed; to define the accuracy and synchronization reference when the coordinator retains its current backbone state and version gateway communication is unconstrained. We then evaluate the for the next communication round. The same uniform updatefull P1–P6 policy landscape under the default heterogeneousinstance aggregation rule is used across all policies, while link profile, showing how participation, buffering, stale-update scheduling, buffering, checkpoint availability, and stale-update handling, and download control create distinct accuracy– admission determine which update instances enter the accepted communication operating points. To separate the effect of set. Only backbone updates enter the aggregation; subjectcommunication-aware coordination from the broader policy specific classifier heads remain gateway-local. landscape, we compare P5 with the matched non-adaptive G. Stale-Aware Backbone Downloads FIFO stale-rejection policy P3 and further decompose the The download rule determines when an online gateway communication-aware policy through a component analysis. We refreshes its cached backbone from the edge coordinator then assess the robustness of the principal P3/P5 contrast across before local training. Under always-download synchroniza- five matched training and gateway-availability realizations. We tion, a selected gateway downloads the coordinator backbone next vary link availability from mild to severe conditions to test whenever the coordinator version is newer than the gateway’s whether the observed trade-offs persist under changing concached version. Under stale-aware synchronization, the gateway nectivity. Finally, we examine subject-level outcomes relative downloads only when its local backbone lag exceeds a fixed to the ideal-link reference, since aggregate accuracy alone can download threshold: conceal large individual gains, losses, and low-accuracy cases. During the gateway-coordinated phase, raw EEG, calibration V r − vir > τd , (13) labels, and subject-specific classifier heads remain at the where V r is the coordinator backbone version at round r and subject-side gateways, and collaboration occurs through sharedbackbone updates; EIB-PH additionally uses pooled Session-1 vir is the backbone version currently cached by gateway i. Stale-aware downloading is instantiated with threshold data for predeployment backbone initialization. The evaluation τd = 1 for all subjects. Because stale-update admission structure follows the research questions: the ideal-link reference allows delayed updates up to τs = 2, the download rule addresses RQ1; the P1–P6 policy landscape, controlled P3/P5 triggers refresh before local training proceeds from a backbone comparison, and communication-aware component analysis that is already at the stale-update boundary. A gateway may address RQ2; and the link-availability sensitivity and subjecttherefore train with a recently cached backbone, but it refreshes level reliability analyses address RQ3. once its local copy is more than one coordinator version A. Experimental Setting behind. This choice reduces unnecessary server-to-gateway Table II summarizes the experimental protocol. We evaluate backbone transfers while limiting the synchronization lag from which subsequent local updates are produced. Together BCICIV-2a and OpenBMI, two public MI-BCI datasets with with Bmax = 3 and τs = 2, the download threshold is complementary cohort sizes and class structures. Both datasets fixed across datasets and subjects so that policy comparisons comprise benchmark MI-EEG recordings collected under reflect coordination behavior rather than threshold tuning. Each controlled laboratory conditions, rather than recordings from policy is therefore treated as a multi-objective operating point rehabilitation patients or home-use deployments. BCICIV-2a over decoding accuracy, client-to-server traffic, server-to-client contains 9 subjects and four motor-imagery classes, whereas traffic, update freshness, and subject-level reliability, rather OpenBMI contains 54 subjects and two classes. Each subject is represented by one subject-side gateway. For both datasets, Sesthan as the optimizer of a single scalar objective. sion 1 supports collaborative backbone learning, and Session 2 V. E XPERIMENTAL E VALUATION supports gateway-side personalization and held-out evaluation. This section evaluates gateway-coordinated federated perEEG trials are represented as four-second epochs sampled at sonalization for motor-imagery brain-computer interfaces (MI- 250 Hz. BCICIV-2a uses 22 EEG channels, and OpenBMI uses BCIs) under heterogeneous gateway-to-edge connectivity. The a selected 20-channel motor-imagery montage. All experiments experiments examine communication policy as an element of use EEGNet as the shared feature-learning backbone, with the learning process rather than as an external transmission dataset-specific input-channel and output-class dimensions. The

EEGNet configuration uses F1 = 8, depth multiplier D = 2, averaged over buffered uploads that reach stale-update admisF2 = 16, temporal kernel length C1 = 125, and dropout sion and reported in communication rounds; and download probability 0.5. Collaborative training uses 120 communication avoidance is measured over synchronization opportunities. rounds, 50 local epochs for each newly produced gateway Subject-level reliability is evaluated using absolute personalized update, Adam optimization with learning rate 10−3 , batch accuracy and accuracy differences relative to the matched idealsize 16, and FedAvg-style uniform averaging of accepted link reference, including median and extreme differences, lowreconstructed backbone states at the update-instance level. accuracy counts, changes of at least five percentage points, and Subject-specific classifier heads are not aggregated. the number of subjects who are both below the task-specific For Session 2 personalization, each gateway adapts only analytical threshold and degraded by at least 5 pp relative to the classifier head using k ∈ {15, 20, 30} labeled calibration the matched ideal-link reference. Statistical inference uses subjects as the unit of analysis. trials per class. The chronological Session 2 split keeps the held-out test set fixed across calibration budgets. BCICIV-2a For each subject, policy, and learning regime, accuracy is uses the first 144 Session 2 trials as the calibration pool and first averaged across the three calibration budgets so that the remaining 144 trials for testing; OpenBMI uses the first 80 repeated measurements from the same subject are not treated Session 2 trials as the calibration pool and the remaining 120 as independent observations. Controlled policy comparisons trials for testing. The Session 2 head is initialized from the are then based on paired subject-level differences. We report subject’s latest learned Session 1 head when available and is the mean paired accuracy difference in percentage points, its trained with learning rate 5 × 10−5 , a maximum of 750 epochs, two-sided 95% Student-t confidence interval, the standardized paired effect size dz , the two-sided paired t-test p-value, and and early-stopping patience of 100. The ideal-link reference removes communication- the Wilcoxon signed-rank p-value. The paired effect size is induced missed uploads and delayed delivery. The default ¯ ∆ heterogeneous-link profile uses high-, moderate-, and lowdz = , s∆ availability probabilities of 0.95/0.70/0.40, with target group ¯ and s∆ are the mean and standard deviation of the fractions of 0.34/0.33/0.33. Communication-aware scheduling where ∆ uses gateway-selection budgets of K = 6 for BCICIV-2a paired subject-level differences, respectively. and K = 40 for OpenBMI. For policies that enable the For the principal P5-versus-P3 comparison, Holm adjustment corresponding mechanisms, the buffer capacity is fixed at is applied across the four paired t-tests defined by the two Bmax = 3, stale-update rejection uses τs = 2 backbone datasets and two learning regimes. The link-availability analysis versions, and stale-aware backbone downloading uses τd = 1 is treated as a sensitivity analysis: inference for the default backbone version. To support controlled policy comparisons, profile is taken from the prespecified primary comparison, while all policies within a given realization use identical data the additional profile-specific comparisons are summarized partitions, Session 2 calibration samples, model initialization, using paired differences and 95% confidence intervals. The gateway-group assignment, and realized availability trace. The Wilcoxon signed-rank tests are reported as complementary complete P1–P6 landscape, component analysis, availability- unadjusted analyses. Given the smaller BCICIV-2a cohort severity study, and subject-level association analysis are (n = 9), its results are interpreted using confidence intervals reported for the primary realization using seed 2026. The and effect sizes alongside the paired tests. The complete principal P3/P5 contrast is additionally evaluated across five P1–P6 operating points across accuracy, traffic, coordinator matched realizations with different training initializations rejection, accepted-update staleness, and buffered-upload delay and gateway-availability traces to determine whether its are reported in Supplementary Table S1. accuracy and communication effects persist beyond the Robustness of the principal P5-versus-P3 comparison is primary realization; the full robustness analysis is reported in assessed across five matched realizations. The primary repeatedSupplementary Section S2. run interval is a crossed hierarchical bootstrap with 10,000 resamples of matched realizations and subjects while preserving B. Metrics and Statistical Methodology the within-subject policy pairing. Across-realization t intervals The primary learning outcome is subject-level decoding ac- and observed ranges are reported descriptively, and each curacy after gateway-side head-only personalization. Accuracy matched realization serves as the replicate unit for communiis reported separately by calibration budget where relevant cation outcomes. Because several stochastic elements of the and averaged across k ∈ {15, 20, 30} for aggregate policy experimental realization vary across the matched replicates, comparisons. Communication is measured as client-to-server the analysis characterizes their combined influence rather than model-update traffic and server-to-client backbone traffic in isolating individual sources of run-to-run variability. decimal MB. Synchronization behavior is characterized by the coordinator-rejected upload rate, mean accepted-update C. Ideal-Link Reference Performance staleness, mean buffered-upload delay, and download avoidance. The ideal-link reference establishes the behavior of the The rejection rate is computed over transmitted uploads; learning pipeline under unconstrained gateway-to-edge synchroaccepted-update staleness is averaged over admitted uploads nization. In this condition, selected gateways synchronize when and reported in backbone versions; buffered-upload delay is required, uploads are delivered without communication-induced

TABLE II E XPERIMENTAL SETTING FOR GATEWAY- COORDINATED FEDERATED MI-BCI PERSONALIZATION . A. Data and learning protocol Evaluation component

Configuration used in the experiments

Datasets and gateways

BCICIV-2a: 9 subjects and four classes; OpenBMI: 54 subjects and two classes. Each subject is modeled as one subject-side gateway. Four-second epochs sampled at 250 Hz. BCICIV-2a uses 22 EEG channels; OpenBMI uses a selected 20-channel motor-imagery montage. Shared EEGNet backbone with dataset-specific input and output dimensions; F1 = 8, D = 2, F2 = 16, C1 = 125, and dropout probability 0.5. Session 1 supports collaborative backbone learning. Session 2 supports gateway-side head personalization and held-out testing. SB-PH uses a commonly initialized backbone for federated learning with subject-specific heads. EIB-PH initializes the backbone through pooled Session 1 pretraining and then follows the same federated-training and head-only personalization protocol. Pooled Session 1 trials from the participating training subjects with a seeded random 80/20 class-stratified trial-wise split; subjects may contribute trials to both subsets, and validation is used for early stopping and model selection rather than unseen-subject evaluation. Adam optimizer; learning rate 10−3 ; maximum 1500 epochs; early-stopping patience 200. 120 communication rounds; 50 local epochs for each newly produced gateway update; Adam optimizer; learning rate 10−3 ; batch size 16. FedAvg-style uniform averaging of accepted reconstructed backbone states at the update-instance level. Subject-specific classifier heads remain gateway-local. k ∈ {15, 20, 30} labeled Session 2 calibration trials per class, selected chronologically from a fixed calibration pool. The last selected trial from each class is reserved for early-stopping validation, and the remaining selected trials are used for head updates. The held-out test set remains fixed across calibration budgets. BCICIV-2a: first 144 trials for the calibration pool and remaining 144 trials for testing. OpenBMI: first 80 trials for the calibration pool and remaining 120 trials for testing. The available backbone is held fixed, and only the subject-specific classifier head is updated. The head is initialized from the latest locally stored Session 1 head, with the common initial head used if no Session 1 head is available; learning rate 5 × 10−5 ; maximum 750 epochs; early-stopping patience 100.

EEG preprocessing EEGNet configuration Session protocol Learning regimes

EIB-PH initialization

Federated training Aggregation rule Calibration sampling

Session 2 split Head personalization

B. Communication and statistical protocol Evaluation component

Configuration used in the experiments

Communication reference

Ideal-link reference with uninterrupted gateway availability, no communication-induced missed or delayed uploads, and immediate backbone synchronization when a newer coordinator version is available. High-, moderate-, and low-availability probabilities of 0.95/0.70/0.40, with target gateway-group fractions of 0.34/0.33/0.33. Gateway-selection budget K = 6 for BCICIV-2a and K = 40 for OpenBMI; FIFO buffer capacity Bmax = 3 updates per gateway; stale-update admission threshold τs = 2 backbone versions; stale-aware download threshold τd = 1 backbone version. Primary analyses use seed 2026. For robustness, the P3/P5 comparison is repeated across five matched realizations with distinct training initializations and gateway-availability traces. Held-out subject-level accuracy and reliability relative to the ideal-link reference; client-to-server model-update and server-to-client backbone traffic in decimal megabytes (1 MB = 106 bytes); coordinator-rejected uploads as a percentage of transmitted uploads; event-weighted mean accepted-update staleness in backbone versions; event-weighted mean buffered-upload delay in communication rounds; and download avoidance as a percentage of synchronization opportunities. Paired tests use subjects as the statistical unit after first averaging each subject’s accuracy across the three calibration budgets.

Default heterogeneous links Communication parameters

Seed and availability traces Metrics and units

Statistical unit

delay, and refreshed backbone states are available without missed downloads. The purpose of this condition is to provide a stable synchronization reference for RQ1 and a subject-specific baseline for the later reliability analysis. It is best interpreted as a reference synchronization condition rather than an accuracy ceiling. Ideal synchronization maximizes the regularity of model exchange, but it does not necessarily produce the highest empirical accuracy for every dataset, regime, or calibration budget. This distinction is important because heterogeneous-link policies also change the learning trajectory. When synchronization is constrained, the coordinator may aggregate a different

sequence of subject updates, admit delayed updates with bounded staleness, reject older updates, or refresh gateway backbones less frequently. These mechanisms can alter the effective optimization path of the shared backbone. In limitedcalibration and subject-heterogeneous EEG learning, such changes can sometimes act as implicit regularization: fewer backbone refreshes may reduce sensitivity to round-to-round update variation, bounded stale updates may smooth or redirect the aggregation trajectory, and stochastic gateway availability may produce small empirical gains for some dataset-regime combinations. Therefore, heterogeneous-link accuracy can occasionally exceed the ideal-link reference, even though the

ideal-link condition remains the cleanest reference for inter- P6 reached 68.44% and 64.51%, respectively. The complete preting communication-constrained operation. Supplementary policy landscape therefore shows that communication efficiency Figure S1 reports mean decoding accuracy across the three and decoding accuracy cannot be ranked independently of the Session 2 calibration budgets for the ideal-link reference and dataset and learning regime. Policies P1–P6 under both learning regimes. The full P1–P6 comparison uses a common training initialOpenBMI shows a clear separation between the two learning ization and a matched gateway-availability trace to support regimes under the ideal-link reference. SB-PH reaches 72.23% controlled policy comparisons. Robustness of the principal mean accuracy, whereas EIB-PH reaches 80.75%. This 8.52 P3/P5 comparison across five matched training and gatewaypercentage-point difference indicates that the pooled Session 1 availability realizations under the default heterogeneous-link representation initialization provides a substantial reference- profile is summarized in Table V; complete replicate-level, setting advantage on this two-class cohort. On BCICIV-2a, subject-level, and communication results are reported in Supthe two regimes are closer: SB-PH reaches 67.77%, while plementary Section S2 (Tables S2–S5). EIB-PH reaches 66.82%. Thus, the value of the representationFigure 2 relates mean subject accuracy to server-to-client initialized regime is dataset-dependent, with a clear ideal-link backbone traffic. Lower downlink traffic does not necessarily gain on OpenBMI and no corresponding ideal-link advantage produce an accuracy-neutral operating point because scheduling, on BCICIV-2a. buffering, stale-update admission, and backbone-download The ideal-link rows also establish the communication ex- decisions alter the sequence, recency, and composition of posure of unconstrained collaboration. BCICIV-2a requires the updates that shape the shared backbone. The resulting 9.30 MB of client-to-server traffic and 9.23 MB of server-to- points therefore represent coupled learning and synchronization client traffic. OpenBMI requires 55.00 MB and 54.56 MB, outcomes rather than bandwidth reductions applied after respectively. These values define the synchronization-cost refer- training. ence against which heterogeneous-link policies are interpreted. Figure 3 decomposes client-to-server and server-to-client Accordingly, the heterogeneous-link policies are interpreted traffic under EIB-PH. P5 and P6 produce the largest reduction as joint learning-and-synchronization operating points that in the server-to-client component, consistent with their stalecouple personalized decoding accuracy with client-to-server aware backbone-download rule. Client-to-server traffic varies update volume, server-to-client backbone synchronization, less because fresh and buffered updates are still transmitted update freshness, and delayed-update handling. This framing when gateways obtain upload opportunities. is necessary because each policy changes not only the amount of communication, but also the timing and composition of the E. Effect of Communication-Aware Synchronization The P3/P5 comparison is the central controlled analysis of backbone updates that shape subsequent personalization. communication-aware synchronization under heterogeneous D. Accuracy and Communication under P1–P6 gateway links. Both policies use FIFO buffering and staleThe P1–P6 comparison examines how gateway scheduling, update rejection, so the offline-update and stale-admission rules offline-update handling, stale-update admission, and backbone- are held fixed. They differ in two synchronization dimensions: download control jointly determine decoding performance and gateway scheduling and backbone-download control. P3 uses communication under the default heterogeneous-link profile. non-adaptive all-gateway scheduling with always-download Supplementary Table S1 reports the complete operating points, synchronization, whereas P5 combines communication-aware including mean subject accuracy and the corresponding com- gateway scheduling with stale-aware backbone downloading. munication and synchronization measurements. The comparison therefore isolates the joint effect of adapting On OpenBMI, EIB-PH achieved higher mean subject accu- gateway scheduling to realized availability and backbone racy than SB-PH under the ideal-link reference and every refresh to version lag while holding the buffering and staleheterogeneous-link policy. Within EIB-PH, mean accuracy update admission rules constant. ranged from 80.13% under P3 to 81.12% under P5, compared Table III shows that P5 substantially reduces server-to-client with 80.75% under the ideal-link reference. P5 also achieved the backbone traffic relative to P3. On OpenBMI, server-to-client highest OpenBMI SB-PH accuracy, 73.02%. P5 and P6 reduced traffic decreases from 37.49 MB to 21.98 MB under both SBserver-to-client backbone traffic to 21.98 MB, compared with PH and EIB-PH, corresponding to a 41.38% reduction and 37.49 MB under P3 and P4 and 54.56 MB under the ideal-link 15.51 MB saved relative to P3. On BCICIV-2a, it decreases reference. These savings were accompanied by greater update from 6.27 MB to 3.68 MB under both regimes, corresponding rejection and accepted-update staleness, reflecting the trade-off to a 41.34% reduction and 2.59 MB saved relative to P3. When introduced by communication-aware synchronization. client-to-server and server-to-client traffic are combined, P5 BCICIV-2a showed a different policy ordering. P4 achieved saves 15.72 MB of total traffic on OpenBMI and 3.17 MB the highest mean accuracy under both SB-PH and EIB-PH, on BCICIV-2a relative to P3. By comparison, client-to-server reaching 68.88% and 69.80%, respectively. Although P5 and traffic decreases by only 0.21 MB on OpenBMI and 0.58 MB P6 reduced server-to-client traffic to 3.68 MB, their accuracy on BCICIV-2a. The dominant communication saving therefore depended more strongly on the personalization regime. P5 occurs in the server-to-client direction through P5’s stale-aware reached 68.44% under EIB-PH and 66.77% under SB-PH, while backbone-download rule.

BCICIV-2a, SB-PH

BCICIV-2a, EIB-PH 72

72 70

S2C reduction: 41.34%

S2C reduction: 41.34%

70

68 68 66 66

Mean accuracy (%)

64 3

4

5

6

7

8

9

10

3

OpenBMI, SB-PH

5

82

S2C reduction: 41.38%

6

7

8

9

10

OpenBMI, EIB-PH

83

74 73

4

S2C reduction: 41.38%

81

72

80

71

79

70 20

30

40

50

60

20

30

40

50

60

Server-to-client backbone traffic (MB) Ideal

P1

P2

P3

P4

P5

P6

P3 to P5

Fig. 2. Mean held-out decoding accuracy versus server-to-client (S2C) backbone traffic under the ideal-link reference and the default heterogeneous-link policies. For each policy, accuracy was first averaged within each subject across the three Session 2 calibration budgets (k ∈ 15, 20, 30 trials per class) and then across subjects (BCICIV-2a, n = 9; OpenBMI, n = 54). The light-gray vertical error bars centered on P3 and P5 denote ±1 standard error across the subject-level averages. In each panel, the black arrow shows the change in the accuracy–downlink operating point from P3 to P5. The adjacent annotation reports P5’s reduction in S2C backbone traffic relative to P3: 41.34% for BCICIV-2a and 41.38% for OpenBMI. Axis ranges are panel-specific; numerical values should be read from the axis ticks, and visual distances should not be compared across panels.

Figure 4 complements the aggregate comparison by showing P3 and P5 across the evaluated Session 2 calibration budgets, with the ideal-link result included as the unconstrained synchronization reference. On OpenBMI, P5 remains above P3 across all three calibration budgets under both learning regimes, with the strongest aggregate result under EIB-PH. On BCICIV-2a, the P3/P5 relationship varies across calibration budgets and learning regimes. The server-to-client communication reduction is therefore consistent, whereas the corresponding accuracy effect depends on the dataset and learning regime. The strongest accuracy result occurs on OpenBMI under EIB-PH. In this setting, P5 increases mean subject accuracy from 80.13% to 81.12% relative to P3. The paired comparison yields a mean difference of +0.99 percentage points, with a 95% confidence interval of [0.34, 1.63] percentage points and Holm-adjusted p = 0.013. Table III summarizes this accuracy–communication result, while Table IV reports the full paired statistical analysis. This is the clearest case in which communication-aware synchronization improves personalized

accuracy while also reducing server-to-client backbone traffic. Under OpenBMI SB-PH, P5 increases mean subject accuracy from 71.88% to 73.02%, with a mean paired difference of +1.14 percentage points. The unadjusted paired t-test gives p = 0.026, but the Holm-adjusted value is p = 0.078, so the improvement does not remain statistically significant after adjustment. On BCICIV-2a, P5 preserves the server-to-client communication reduction, but the accuracy differences are not statistically supported. Under EIB-PH, P5 and P3 are nearly identical at 68.44% and 68.47%, respectively, with Holm-adjusted p = 0.981. Under SB-PH, P5 is lower than P3, at 66.77% versus 68.13%, but the paired confidence interval includes zero, and the Holm-adjusted value is p = 0.906. Taken together, the primary P3/P5 comparison shows that P5 reduced server-to-client backbone traffic by approximately 41% on both datasets and also reduced total communication, with the only Holm-adjusted accuracy difference observed on OpenBMI under EIB-PH. To assess whether this operating point depended on the

BCICIV-2a Model communication traffic (MB)

20

120

18.5 15.3

15

15.3

OpenBMI 109.6

Client-to-server Server-to-client

100

90.3

14.2

12.6

12.1

11.1

90.3

84.5

75.1

80

74.6

68.8

60

10

40 5

20

0

0 Ideal

P1

P2

P3

P4

P5

P6

Ideal

P1

P2

P3

P4

P5

P6

Fig. 3. Client-to-server (C2S) and server-to-client (S2C) model communication traffic for the ideal-link reference and the default heterogeneous-link policies. The plotted values are from EIB-PH; the corresponding SB-PH values are identical at the reported MB precision, indicating that the communication pattern shown is not specific to EIB-PH. The total bar height denotes aggregate model traffic, and stale-aware backbone downloading primarily reduces the S2C component. TABLE III C ONTROLLED P3/P5 COMPARISON UNDER THE DEFAULT HETEROGENEOUS - LINK CONDITION .

A. Observed operating points Mean subject accuracy (%)

C2S traffic (MB)

S2C traffic (MB)

Rejected uploads Accepted-update Avoided downloads (%) staleness (versions) (%)

Dataset

Regime Policy

BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a

SB-PH SB-PH EIB-PH EIB-PH

P3 P5 P3 P5

68.13 66.77 68.47 68.44

9.02 8.44 9.02 8.44

6.27 3.68 6.27 3.68

19.20 24.59 19.20 24.59

0.27 0.52 0.27 0.52

0.00 35.92 0.00 35.92

OpenBMI OpenBMI OpenBMI OpenBMI

SB-PH SB-PH EIB-PH EIB-PH

P3 P5 P3 P5

71.88 73.02 80.13 81.12

52.83 52.62 52.83 52.62

37.49 21.98 37.49 21.98

18.51 22.03 18.51 22.03

0.25 0.55 0.25 0.55

0.00 41.55 0.00 41.55

B. Paired accuracy effects and communication savings Paired difference (P5–P3, pp)

95% CI (pp)

Holm-adjusted p-value

S2C traffic saved (MB)

S2C reduction (%)

Total traffic saved (MB)

Regime

n

BCICIV-2a SB-PH BCICIV-2a EIB-PH

9 9

−1.36 [−5.35, 2.62] −0.03 [−2.47, 2.42]

0.906 0.981

2.59 2.59

41.34 41.34

3.17 3.17

OpenBMI OpenBMI

54 54

+1.14 +0.99

0.078 0.013

15.51 15.51

41.38 41.38

15.72 15.72

Dataset

SB-PH EIB-PH

[0.14, 2.14] [0.34, 1.63]

Notes: Paired accuracy differences are computed as P5 minus P3. The 95% CIs are two-sided Student-t intervals based on paired subject-level differences. Holm adjustment is applied to the four paired t-test p-values defined by the two datasets and two learning regimes. S2C traffic saved and total traffic saved report the communication savings achieved by P5 relative to P3 and are computed as P3 minus P5; S2C reduction is likewise measured relative to P3. Total traffic is C2S plus S2C traffic. Rejected uploads are coordinator-rejected transmitted uploads; avoided downloads are normalized by synchronization opportunities. C2S and S2C denote client-to-server and server-to-client, respectively.

primary controlled realization, the P3/P5 comparison was were small and realization-dependent, and all four crossed repeated across five matched training and gateway-availability hierarchical-bootstrap 95% confidence intervals included zero. realizations under the default heterogeneous-link profile. Ta- The repeated analysis therefore supports the robustness of ble V summarizes the principal repeated-analysis outcomes; the communication saving, but does not establish a consistent complete replicate-level, subject-level, and communication accuracy effect beyond the primary realization. results are reported in Supplementary Section S2 (Tables S2– S5). F. Component Analysis of P5 P5 reduced server-to-client backbone traffic in every matched realization, with mean reductions of 42.23% on BCICIV-2a and 41.98% on OpenBMI. In contrast, the accuracy differences

The controlled P3/P5 comparison establishes the net effect of the full communication-aware policy. However, P5 combines two coordination mechanisms that can influence the operating

BCICIV-2a, SB-PH

Accuracy (%)

70

BCICIV-2a, EIB-PH 70

68

68

66

66

64

64

Accuracy (%)

OpenBMI, SB-PH

OpenBMI, EIB-PH 81.25

73.0

81.00 80.75

72.5

80.50 72.0

80.25 80.00 15

20

30

15

Calibration trials per class

20

30

Calibration trials per class

Ideal-link reference

P3

P5

Fig. 4. Mean subject accuracy for P3 and P5 across Session 2 calibration budgets. The ideal-link result is included as the unconstrained synchronization reference for the corresponding dataset and learning regime. TABLE IV S UBJECT- LEVEL PAIRED ACCURACY COMPARISON OF P5 AND P3 UNDER THE DEFAULT HETEROGENEOUS - LINK CONDITION .

Dataset

Regime

n

Mean paired difference (pp)

95% Student-t CI (pp)

Paired effect size (dz )

Paired t-test p-value

Wilcoxon signed-rank p-value

Holm-adjusted t-test p-value

BCICIV-2a BCICIV-2a

SB-PH EIB-PH

9 9

-1.36 -0.03

[−5.35, 2.62] [−2.47, 2.42]

-0.26 -0.01

0.453 0.981

0.496 1.000

0.906 0.981

OpenBMI OpenBMI

SB-PH EIB-PH

54 54

1.14 0.99

[0.14, 2.14] [0.34, 1.63]

0.31 0.42

0.026 0.003

0.016 0.001

0.078 0.013

Notes: Accuracy differences are computed as P5 minus P3 and reported in percentage points. The 95% CIs are two-sided Student-t intervals based on paired subject-level accuracy differences. Holm adjustment is applied across the four paired t-tests defined by the two datasets and two learning regimes; Wilcoxon signed-rank tests are complementary unadjusted analyses.

point in different ways: online-priority scheduling and staleaware backbone downloading. We therefore decompose P5 on OpenBMI under EIB-PH, the setting with the largest subject cohort and the strongest controlled improvement of P5 relative to P3. This analysis separates online-first gateway selection from backbone-download control while keeping FIFO buffering and stale-update rejection fixed.

of only 0.56%. The downlink saving observed under P5 is therefore primarily attributable to its backbone-download rule.

The accuracy results show a complementary contribution from gateway scheduling. Online-random scheduling increases mean subject accuracy from 80.13% to 80.37%, whereas online-priority scheduling reaches 80.67%. The higher accuracy under online-priority scheduling indicates that prioritizing Table VI shows that stale-aware backbone downloading gateways according to synchronization need contributes beyond produces the principal server-to-client traffic reduction. Adding restricting participation to gateways that are currently online. stale-aware downloading without online-priority scheduling Stale-aware downloading alone reaches 80.43%, combining the reduces server-to-client backbone traffic from 37.49 MB to full downlink reduction with a smaller accuracy increase. Full 21.98 MB, matching the 41.38% reduction obtained by full P5. P5 attains the strongest component setting, reaching 81.12% In contrast, online-random and online-priority scheduling retain while retaining the same 21.98 MB of server-to-client backbone server-to-client traffic near the P3 reference, with reductions traffic as the stale-aware-download-only variant.

constrained, while accepted-update staleness remains higher under P5. These results confirm that severe availability creates a more delayed and more selective update stream for both policies. Mean accuracy difference Mean S2C reduction Across all availability profiles, P5 maintains a lower server-toCondition (P5–P3, pp) (%) client synchronization cost than P3. On OpenBMI, P5 reduces [95% CI] [observed range] server-to-client backbone traffic from 44.36 MB to 24.39 MB BCICIV-2a, SB-PH +0.80 [−1.04, 3.08] 42.23 [41.59, 42.76] under mild availability, from 37.49 MB to 21.98 MB under the BCICIV-2a, EIB-PH +0.88 [−1.16, 2.97] 42.23 [41.59, 42.76] default profile, and from 28.99 MB to 17.81 MB under severe OpenBMI, SB-PH −0.09 [−0.78, 0.67] 41.98 [41.47, 42.47] OpenBMI, EIB-PH −0.07 [−0.53, 0.39] 41.98 [41.47, 42.47] availability. On BCICIV-2a, the corresponding reductions are from 7.46 MB to 4.07 MB, from 6.27 MB to 3.68 MB, and Notes: Accuracy differences are computed as P5 minus P3 and reported in percentage points. The 95% confidence intervals are obtained using a from 4.86 MB to 2.99 MB. Thus, the downlink-saving effect crossed hierarchical bootstrap across matched realizations and subjects. S2C of P5 persists across the full range of evaluated link conditions. reduction is measured relative to P3; the bracketed values report the observed The accuracy effect is less stable than the communication minimum and maximum across the five realizations. S2C values are common to SB-PH and EIB-PH at the reported precision. S2C denotes server-to-client. effect. On OpenBMI, P5 is nearly tied with P3 under mild availability, 80.61% versus 80.53%; it is higher under the default profile, 81.12% versus 80.13%; and it is lower unThese results clarify the operating point achieved by P5 on der severe availability, 80.73% versus 81.19%. The defaultOpenBMI under EIB-PH. Stale-aware backbone downloading profile improvement corresponds to the prespecified primary controls server-to-client synchronization cost, while onlinecomparison reported in Table IV (Holm-adjusted p = 0.013), priority scheduling changes which available gateway updates while the 95% confidence intervals for the mild- and severeshape the shared backbone. Their combination yields the profile comparisons include zero. On BCICIV-2a, P5 is higher highest observed accuracy among the component variants than P3 under mild availability, 70.06% versus 68.29%, and while preserving the full downlink reduction, reinforcing the nearly matched under the default and severe profiles; the 95% conclusion that communication-aware coordination influences confidence intervals for all three profile-specific comparisons both synchronization cost and the learning trajectory of include zero. personalized MI-BCI. These results refine the P3/P5 conclusion. P5 consistently reduces server-to-client backbone synchronization as link G. Sensitivity to Heterogeneous Link Availability availability changes, but its accuracy effect depends on the Gateway availability can change with wireless conditions, availability profile and dataset. More constrained availability user mobility, device power state, and network congestion. also increases coordinator-rejected uploads and buffered-upload We therefore test whether the P3/P5 conclusions remain delay, while accepted-update staleness remains dependent stable as link availability changes from mild to default to on the policy and availability profile. The same policy can severe conditions. The sensitivity analysis uses EIB-PH and therefore operate on a more delayed and more selective update compares P3 with P5 under three online-probability pro- stream as connectivity deteriorates. Thus, communication-aware files: mild 0.98/0.85/0.60, default 0.95/0.70/0.40, and severe synchronization should not be treated as a universal default 0.90/0.50/0.20 for high-, moderate-, and low-availability without validation on the intended user cohort, personalization gateways. P3 and P5 are used because they share FIFO regime, and gateway-availability conditions. buffering and stale-update rejection, allowing the analysis to vary gateway scheduling and backbone-download control while H. Subject-Level Reliability holding delayed-update handling fixed. Aggregate accuracy is necessary but insufficient for MI-BCI Figure 5 summarizes the principal accuracy–communication deployment. A communication policy can improve the cohort trends across the availability profiles, while Supplementary mean while leaving some users with degraded or low-accuracy Table S6 reports the complete accuracy, communication, and personalized decoders. We therefore compare each subject’s synchronization results. Coordinator-rejected upload rates and heterogeneous-link accuracy with that subject’s matched idealbuffered-upload delays increase as the profile moves from mild link reference under EIB-PH and summarize the resulting to severe. On OpenBMI, the coordinator-rejected upload rate reliability profiles in Table VII. For threshold-based reporting, under P3 rises from 8.63% to 27.52%, while mean buffered- we use 60% accuracy for the four-class BCICIV-2a task and upload delay increases from 1.33 to 1.60 rounds. Under P5, 70% accuracy for the binary OpenBMI task. These thresholds the corresponding values rise from 13.60% to 30.64% and are analytical markers rather than clinical usability criteria. from 1.33 to 1.59 rounds. Mean accepted-update staleness They identify low-accuracy operating regions while accounting remains higher under P5 than under P3 across all three profiles, for the different class structures and expected performance ranging from 0.51 to 0.55 backbone versions under P5 and ranges of the two tasks. from 0.21 to 0.26 backbone versions under P3. BCICIV-2a OpenBMI illustrates why subject-level reliability must be follows the same overall pattern: coordinator rejection and examined alongside aggregate accuracy. Under EIB-PH, P5 has buffered-upload delay increase as availability becomes more a mean subject-level accuracy difference of +0.38 percentage TABLE V ROBUSTNESS OF THE P3/P5 COMPARISON ACROSS FIVE MATCHED TRAINING AND GATEWAY- AVAILABILITY REALIZATIONS UNDER THE DEFAULT HETEROGENEOUS - LINK PROFILE .

TABLE VI C ONTROLLED COMPONENT ANALYSIS OF P5 ON O PEN BMI UNDER EIB-PH AND THE DEFAULT HETEROGENEOUS - LINK CONDITION . A. Controlled component variants Variant

Gateway-selection rule

Backbone-download rule

P3 reference Online-random selection

Fixed all-gateway selection Random selection among currently online gateways Online-first priority selection based on version lag, contribution recency, and pending-update state Fixed all-gateway selection Online-first priority selection based on version lag, contribution recency, and pending-update state

Always download the current backbone Always download the current backbone

Online-priority selection Stale-aware download only Full P5

Always download the current backbone Stale-aware backbone download Stale-aware backbone download

B. Accuracy and communication outcomes Variant P3 reference Online-random selection Online-priority selection Stale-aware download only Full P5

Mean subject accuracy (%)

Accuracy difference from P3 (pp)

C2S update traffic (MB)

S2C backbone traffic (MB)

S2C reduction from P3 (%)

Coordinator-rejected uploads (%)

Mean accepted-update staleness (versions)

80.13 80.37 80.67 80.43 81.12

0.00 +0.23 +0.53 +0.30 +0.99

52.83 52.61 52.62 52.83 52.62

37.49 37.27 37.27 21.98 21.98

0.00 0.56 0.56 41.38 41.38

18.51 18.66 18.58 22.01 22.03

0.25 0.25 0.25 0.55 0.55

Notes: Panel A defines the controlled component variants, and Panel B reports their corresponding accuracy and communication outcomes. FIFO buffering and stale-drop admission are held fixed across all variants so that the effects of gateway selection and backbone-download control can be compared. Accuracy differences and S2C traffic reductions are reported relative to the P3 reference. C2S and S2C denote client-to-server and server-to-client, respectively.

points relative to the ideal-link reference, whereas P3 has a results show that the policy with the strongest cohort mean is mean difference of -0.61 percentage points. P5 also increases not necessarily the policy with the most favorable subject-level the number of subjects with an accuracy gain of at least 5 pp tail behavior. We next examine whether subject-level accuracy differences from zero under P3 to four. However, P5 still has two subjects with an accuracy decrease of at least 5 pp, a minimum subject- covary with baseline decoding strength or communication level difference of -7.22 percentage points, and 12 subjects exposure. Supplementary Table S7 reports descriptive Spearman below the 70% analysis threshold. Only one subject under P5 rank correlations together with 95% bootstrap confidence is both below this threshold and degraded by at least 5 pp intervals. On BCICIV-2a, ideal-link accuracy produced the relative to the matched ideal-link reference. This distinction largest point estimate under both P3 and P5 (ρ = −0.68), shows that low absolute accuracy and substantial degradation but the corresponding confidence intervals were wide and relative to the ideal-link reference capture related but distinct included zero. On OpenBMI, all point estimates were small, aspects of subject-level reliability. Its subject-level difference and all reported confidence intervals likewise included zero. range is 13.06 percentage points. Thus, the policy with the Across datasets and policies, no recorded communicationstrongest aggregate OpenBMI result still leaves a nontrivial exposure quantity showed a consistent association with subjectlevel accuracy differences. These exploratory correlations low-accuracy tail. therefore do not support attributing subject-level vulnerability to BCICIV-2a shows an even sharper separation between cohort- rejected uploads, accepted-update staleness, buffering, gateway level improvement and subject-level reliability. Under EIB- availability, or avoided backbone downloads individually. PH, P4 has the largest mean subject-level difference, +2.98 These reliability results have direct significance for MI-BCI percentage points, and five subjects improve by at least 5 pp. deployment. For users already operating in a low-accuracy At the same time, its minimum subject-level difference is range, an additional policy-induced decline can make motor-14.12 percentage points, and its difference range is 22.45 intent feedback unreliable and disrupt continued home neupercentage points. P5 exhibits a similar tail pattern, with a mean rorehabilitation. Operationally, a persistent threshold violation difference of +1.62 percentage points, a minimum difference or a large decline from the subject’s reference performance of -12.04 percentage points, and a difference range of 21.30 could prompt the collection of a small additional calibration percentage points. P6 has the same mean difference as P5, set, a temporary shift to more conservative synchronization +1.62 percentage points, but a substantially narrower difference that favors fresher backbone states, or rollback to the most range of 11.34 percentage points and a minimum difference recent locally validated decoder state. Monitoring both cohortof -5.56 percentage points. Notably, under every P1–P6 policy, level and subject-level outcomes is therefore essential in home no BCICIV-2a subject was simultaneously below the 60% use, where frequent clinician-supervised recalibration may be analytical threshold and degraded by at least 5 pp relative impractical. to the matched ideal-link reference. In other words, for each VI. D ISCUSSION AND D EPLOYMENT I MPLICATIONS policy, the subjects below the 60% analytical threshold were not the same subjects as those who experienced a decline of at The results position federated MI-BCI personalization as least 5 pp relative to the matched ideal-link reference. These a deployment-coupled learning problem rather than a model-

BCICIV-2a, EIB-PH Mild

Mean accuracy (%)

70.0

81.2

Default

Severe

80.8 69.0

Severe

80.4

Default Mild Severe

3

Severe

4

5

Default

80.0 6

7

Server-to-client communication (MB)

15 P3

20

25

30

35

40

Server-to-client communication (MB)

P5

P5 better

3.0

P5 better

1.0

0.0 -3.0

Mild

Mild

Default

68.0

P5 accuracy gain over P3 (percentage points)

OpenBMI, EIB-PH

0.0 P3 better

Mild

-1.0 Default (primary)

Severe

Gateway-availability profile

P3 better

Mild

Default (primary)

Severe

Gateway-availability profile

Fig. 5. Sensitivity of the controlled P3/P5 comparison to mild, default, and severe heterogeneous gateway availability under EIB-PH. Both policies use FIFO buffering and stale-update rejection, isolating the combined effect of communication-aware scheduling and stale-aware backbone downloading. Accuracy is averaged within each subject across k ∈ 15, 20, 30 calibration trials per class and then across subjects (BCICIV-2a, n = 9; OpenBMI, n = 54). The upper panels show mean held-out accuracy versus server-to-client backbone traffic, with arrows tracing mild to default to severe availability; P5 uses less traffic under every profile. The lower panels show mean paired P5-minus-P3 accuracy differences with two-sided Student’s t 95% confidence intervals; positive values favor P5, negative values favor P3, and the dashed line denotes no difference. The light-gray bands behind the lower-panel default columns and the black rings around default markers in both rows identify the prespecified primary profile. Only the OpenBMI default-profile improvement is statistically supported (Holm-adjusted p = 0.013); the remaining comparisons are not statistically resolved.

only problem. In a gateway-coordinated system, communi- experimental variables and relates them to the personalized cation policy determines which subject updates reach the decoder ultimately available to each subject. coordinator, when delayed updates are admitted, and which NEXUS-MI also differs in emphasis from communicationbackbone version is available for subsequent local training and aware federated-learning systems such as Oort and FedBuff, personalization. Synchronization behavior therefore influences which study participant selection and buffered asynchronous the trajectory of the shared representation as well as the comaggregation primarily through time-to-accuracy, scalability, munication required to maintain it. The differences observed convergence, and aggregate model utility [40], [42]. The present across P1–P6 show that gateway coordination cannot be treated framework applies related coordination principles to MI-BCI, as a bandwidth-control layer applied independently of learning. where communication efficiency must be interpreted together with limited-calibration personalization and subject-level reliaThis perspective extends prior federated EEG and MI-BCI bility. The component analysis illustrates this distinction: staleresearch. Existing studies demonstrate that federated learning aware backbone downloading produces the principal reduction can exploit cross-subject information, address non-IID subject in server-to-client traffic, while online-priority scheduling data, and support collaboration across heterogeneous EEG changes which available subject updates shape the shared datasets without centralizing the underlying recordings [10]– backbone. Their combination affects both synchronization cost [12]. Their principal emphasis is the learning method and and the subsequent personalization outcome. its resulting classification performance under a prescribed federation process. NEXUS-MI addresses a complementary The dataset-dependent findings further show why no single deployment question: whether later-session personalization communication policy should be assumed to transfer uniformly remains effective when the shared representation is maintained across MI-BCI cohorts. OpenBMI provides a larger 54-subject through intermittent and policy-controlled gateway synchroniza- cohort and a binary motor-imagery task, whereas BCICIV-2a tion. It therefore makes gateway scheduling, offline buffering, contains nine subjects and requires four-class discrimination. stale-update admission, and backbone-download control explicit The larger OpenBMI cohort provides a broader basis for

TABLE VII S UBJECT- LEVEL RELIABILITY UNDER EIB-PH RELATIVE TO THE IDEAL - LINK REFERENCE . A. Subject-level accuracy differences relative to the ideal-link reference Mean accuracy difference (pp)

Median accuracy difference (pp)

Minimum subject difference (pp)

Maximum subject difference (pp)

Difference range (max–min, pp)

Dataset

Policy

BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a

P1 P2 P3 P4 P5 P6

1.41 1.13 1.65 2.98 1.62 1.62

1.62 1.39 3.94 5.56 3.47 3.47

-6.71 -3.47 -14.12 -14.12 -12.04 -5.56

7.18 6.48 10.42 8.33 9.26 5.79

13.89 9.95 24.54 22.45 21.30 11.34

OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI

P1 P2 P3 P4 P5 P6

0.06 -0.32 -0.61 -0.53 0.38 -0.10

0.00 0.00 -0.28 -0.56 0.28 -0.14

-6.67 -7.22 -8.61 -9.17 -7.22 -5.83

5.83 5.00 3.06 4.17 5.83 5.28

12.50 12.22 11.67 13.33 13.06 11.11

B. Absolute accuracy and subject-level reliability counts Minimum subject accuracy (%)

Subjects below threshold (n)

Subjects with ≥5-pp accuracy decrease (n)

Subjects with ≥5-pp accuracy increase (n)

Below threshold with ≥5-pp decrease (n)

P1 P2 P3 P4 P5 P6

54.86 49.54 59.49 57.41 54.17 52.08

3 3 3 1 2 3

1 0 1 1 1 1

2 1 3 5 2 2

0 0 0 0 0 0

P1 P2 P3 P4 P5 P6

53.06 53.61 52.50 53.61 53.33 53.06

12 10 10 12 12 12

2 4 3 3 2 1

1 1 0 0 4 1

1 2 1 3 1 1

Dataset

Policy

BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI

Notes: Panel A summarizes subject-level accuracy differences relative to the matched ideal-link reference, computed as policy minus ideal-link accuracy. Panel B reports absolute accuracy and subject-level reliability counts. The analytical low-accuracy thresholds are 60% for BCICIV-2a and 70% for OpenBMI; subjects exactly at the threshold are not counted as below threshold. Large decreases and increases include changes of exactly 5 pp. The joint count includes subjects who are both below the corresponding threshold and at least 5 pp below their matched ideal-link accuracy.

pooled Session 1 representation initialization and collaborative backbone learning, which is consistent with the clear EIB-PH advantage and the comparatively stable P3/P5 behavior observed on that dataset. In BCICIV-2a, each subject contributes a larger share of the collaborative update stream, and the fourclass personalized decision problem is more demanding. These characteristics are consistent with the greater subject-level sensitivity observed under constrained synchronization. More broadly, the contrast reflects the established dependence of EEG transfer and MI-BCI performance on subject, session, and task heterogeneity [5], [6]. Communication-policy conclusions must therefore be validated for the intended cohort, learning regime, and calibration protocol rather than inferred from a single benchmark.

differences were small and realization-dependent. This stable communication saving must therefore be interpreted alongside the subject-level outcomes concealed by the cohort mean. Importantly, this operating point is achieved without transferring raw EEG trials, calibration labels, or personalized classifier heads during the gateway-coordinated phase. These remain at the subject-side gateway, while collaboration is limited to backbone-related updates, refreshed backbone states, and synchronization metadata.

Subject-level reliability is particularly important for assistive and rehabilitation-oriented MI-BCI because the decoder determines the feedback experienced by each user. A cohort-level improvement does not compensate for a user whose personalized decoder remains inaccurate or deteriorates The evaluation also shows that communication efficiency substantially under constrained synchronization. For users must be interpreted together with reliability. Reducing server- already operating in a low-accuracy range, an additional policyto-client backbone synchronization is valuable for wearable, induced decline can make motor-intent feedback unreliable and edge-enabled, and home-oriented MI-BCI, where gateway disrupt continued home neurorehabilitation [6]. Operationally, connectivity and bandwidth may be limited [3], [30]. However, a persistent threshold violation or a large decline from the lower traffic alone does not establish a suitable operating point. subject’s reference performance could prompt the collection The approximately 41% reduction observed in the primary of a small additional calibration set, a temporary shift to P5-versus-P3 comparison remained approximately 42% across more conservative synchronization that favors fresher backbone five matched realizations, whereas the corresponding accuracy states, a user-specific backbone refresh, or rollback to the

BCICIV-2a 10 5 0

−10

4

5

1

7

9

6

8

3

−15 2

Accuracy change vs. ideal link (pp)

−5

OpenBMI 10 5 0 −5 −10

37

45

5

22

42

13

12

39

50

38

9

53

43

20

19

16

30

52

24

11

2

23

35

40

34

46

33

49

26

28

44

25

6

21

4

8

36

3

31

17

15

7

32

41

18

1

48

29

14

27

10

51

54

47

−15

Subject ID (sorted by P5-P3 change)

P3

P5

Fig. 6. Subject-level accuracy changes under EIB-PH relative to each subject’s matched ideal-link reference, averaged across k ∈ 15, 20, 30 calibration trials per class (BCICIV-2a, n = 9; OpenBMI, n = 54). Each vertical pair connects the P3 and P5 changes for one subject, with subjects sorted by the P5-minus-P3 difference; dotted lines mark changes of ±5 percentage points. These values are relative changes, not absolute accuracies: a subject may improve relative to ideal link while still have low absolute accuracy. Table VII reports absolute accuracy and counts below the dataset-specific analysis thresholds.

most recent locally validated decoder state. These responses are especially relevant in home use, where frequent cliniciansupervised recalibration may be impractical [3].

tion, longitudinal EEG variability, clinician interaction, or network dynamics of a live wearable system used in the home. Connectivity is represented through fixed Bernoulli availability profiles and fixed policy parameters, which provide reproducible operating conditions but do not reproduce bursty outages, correlated link failures, or time-varying bandwidth in operational networks.

Table VIII summarizes the resulting design guidance. The broader implication is that future MI-BCI federated learning studies should report communication behavior and per-user reliability alongside decoding accuracy. Client-to-server traffic, server-to-client synchronization, update rejection, staleness, The principal P3/P5 comparison was repeated across five buffering, and download avoidance describe the system-level matched training and gateway-availability realizations, showing operating point. Subject-level deviations, threshold violations, that the approximately 42% reduction in server-to-client the co-occurrence of low absolute accuracy and large degrada- backbone traffic persists beyond the primary controlled retion from the ideal-link reference, and worst-subject behavior alization. The associated accuracy differences were small and describe whether that operating point remains reliable for realization-dependent, and all crossed hierarchical-bootstrap individual users. Reporting these dimensions together provides 95% confidence intervals included zero. The repeated analysis a more deployment-relevant standard for evaluating federated was limited to P3 and P5 under the default availability profile; personalization in MI-BCI. the broader P1–P6 policy landscape, component analysis, The scope of these findings is limited to an offline, replay- availability-severity study, and subject-level association analysis based evaluation using two public session-based MI-BCI remain based on the primary realization. Because several datasets and a one-subject/one-gateway abstraction. Both stochastic elements of the experimental realization vary across BCICIV-2a and OpenBMI comprise recordings from healthy the matched replicates, the repeated analysis evaluates their participants collected under controlled laboratory conditions; combined influence rather than isolating individual sources of the present evaluation therefore does not establish performance run-to-run variability. Although NEXUS-MI reduced modelfor stroke survivors or other rehabilitation populations. Al- payload traffic, the experiments did not evaluate its end-to-end though this design supports controlled analysis of gateway deployment cost. Under the evaluated configuration of 120 comsynchronization, it does not capture the behavioral adapta- munication rounds and 50 local epochs per generated update,

TABLE VIII D ESIGN GUIDANCE FOR EDGE - ENABLED MI-BCI DEPLOYMENT. Empirical lesson

Deployment guidance

A constrained-link result is meaningful only when interpreted relative to an ideal-link reference for the same learning protocol.

Establish an unconstrained synchronization reference before deploying communication controls. Report how much accuracy, subject-level reliability, and synchronization cost are retained when gateway links become intermittent or selective. Treat gateway coordination as a control plane for learning. Client selection, buffering, stale-update admission, and backbone-refresh timing should be tuned jointly rather than optimized as independent communication knobs. Avoid unconditional backbone downloads. Version-aware and staleness-aware refresh rules are practical mechanisms for reducing edge-to-gateway traffic while allowing gateways to continue local personalization from cached model states. Do not deploy a bandwidth-saving scheduler as a universal default. Accept a communication policy only after validating that it preserves cohort-level accuracy and does not introduce unacceptable subject-level degradation on the target population. Monitor per-user reliability in addition to cohort means. Worst-subject accuracy, threshold violations, large reference-relative declines, and cases where low accuracy coincides with a large decline should inform conservative synchronization, additional calibration, or rollback. Instrument gateways and the edge coordinator to log communication exposure together with decoding outcomes. Use these measures to support operational monitoring, but not as standalone explanations or predictors of subject-level vulnerability. Keep raw EEG and subject-specific classifier heads at the gateway; exchange only backbone-related parameters and synchronization metadata.

Communication policy changes the learning process, not only the number of transmitted bytes. Server-to-client backbone synchronization is a major controllable cost in gateway-coordinated personalization. The accuracy effect of communication-aware synchronization is not uniform across datasets, regimes, and calibration budgets. Aggregate accuracy can mask individual users who experience large losses or remain below a useful operating threshold.

Dropped uploads, update staleness, buffering, and avoided downloads are useful exposure signals, but none showed a consistent association with subject-level accuracy differences. Collaborative personalization can maintain raw-EEG data locality.

Note: The empirical observations summarized here are limited to the evaluated BCICIV-2a and OpenBMI datasets and the modeled heterogeneous gateway-to-edge communication conditions.

gateway computation time, wall-clock synchronization latency, efficiency should therefore not be interpreted as evidence that energy consumption, and network-protocol overhead were not personalized decoding performance is preserved for every user. These findings support a deployment-centered evaluation measured. The reported reductions therefore demonstrate modelpayload communication savings under controlled offline replay, approach in which personalized accuracy is considered jointly but do not yet establish real-time or energy-efficient operation with bidirectional communication, update freshness, linkon physical gateways. Future work should profile gateway availability sensitivity, and subject-level reliability. A suitable computation time, wall-clock synchronization latency, energy coordination strategy is not simply the one that minimizes consumption, and complete network overhead on representative traffic; it must also maintain acceptable performance for gateway hardware. It should also extend the repeated evaluation individual users as connectivity and EEG conditions vary. The evidence remains limited to offline replay of public to the full policy space and broader connectivity conditions, incorporate measured traces from representative home networks, laboratory datasets, predefined Bernoulli gateway-availability and validate the framework longitudinally in rehabilitation profiles with fixed policy parameters, and five matched repopulations. Adaptive per-user synchronization is another alizations of the principal comparison. Future work should important direction, with gateway scheduling, backbone-refresh validate NEXUS-MI longitudinally under measured network frequency, and stale-update handling responding to each user’s conditions and evolving EEG distributions, including bursty recent decoder stability, calibration need, and observed link and correlated outages. It should also develop user-specific synchronization that responds jointly to availability, version condition rather than remaining fixed across the cohort. lag, calibration need, and decoder stability, with safeguards such as additional calibration or rollback when reliability VII. C ONCLUSION deteriorates. These steps would move communication-aware This study shows that gateway synchronization is not merely federated personalization toward dependable edge-enabled and a communication-layer concern in federated motor-imagery home-oriented deployment. BCI personalization; it helps determine the shared representaAUTHOR C ONTRIBUTIONS tion from which each subject’s decoder is personalized. Across Daniel Adu Worae: Conceptualization; Methodology; Softthe six evaluated policies, no strategy is uniformly dominant ware; Investigation; Formal analysis; Writing - original draft; across datasets and personalization regimes. The principal Writing - review & editing. controlled comparison nevertheless demonstrates a consistent Aarthy Nagarajan: Conceptualization; Methodology; Supersystems benefit: communication-aware coordination reduces vision; Writing - review & editing. server-to-client backbone traffic by approximately 42% across F UNDING five matched realizations. The corresponding accuracy effect is not reproducible, however, and cohort averages conceal subjectThis work received institutional research support associated level losses and low-accuracy cases. Improved communication with the Melchor Visiting Assistant Professorship and the Lucy

Family Institute for Data and Society at the University of Notre Dame. No specific external grant supported this work. ACKNOWLEDGMENTS The authors thank Professor Nitesh Chawla for his valuable feedback on this work. E THICS S TATEMENT The present study analyzed existing publicly available, deidentified EEG data from BCICIV-2a and OpenBMI. No participants were recruited and no new human-subject data were collected for this secondary analysis. No additional institutional ethics review was required for this secondary analysis of publicly available, de-identified data. For OpenBMI, the original study was reviewed and approved by the Korea University Institutional Review Board (1040548-KUIRB-16159-A-2), and written informed consent was obtained from all participants [9]. BCICIV-2a was analyzed as the publicly released, de-identified BCI Competition IV dataset; the public dataset description used for this secondary analysis does not provide an ethics-committee reference number [13], [48]. The manuscript contains no identifiable participant information, images, or videos. D ECLARATION OF C OMPETING I NTEREST The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. DATA AND C ODE AVAILABILITY The datasets analyzed in this study are publicly available from their original repositories. BCICIV-2a is available through the BCI Competition IV and BNCI Horizon 2020 repositories as Dataset 001-2014 [48], and OpenBMI is available through GigaDB under DOI 10.5524/100542 [49]. The NEXUS-MI code, experiment outputs, and supporting reproducibility materials are publicly available at https://github.com/WadElla/ NEXUS-MI and are permanently archived on Zenodo under DOI 10.5281/zenodo.22074301 [50]. The reference environment used for the repeated experimental runs was Linux 6.8 (glibc 2.35) with Python 3.10.19, PyTorch 2.7.1, CUDA 11.8, NumPy 2.2.5, and an NVIDIA CUDA-capable device. The archived software supports Python 3.10 or later, and complete dependency and execution requirements are documented with the software release. NEXUS-MI is distributed under the MIT License. D ECLARATION OF G ENERATIVE AI AND AI-A SSISTED T ECHNOLOGIES IN M ANUSCRIPT P REPARATION During the preparation of this work, the authors used ChatGPT (OpenAI) to assist with language refinement, code development, debugging, and statistical analysis of authorgenerated results. The authors reviewed and edited all AIassisted outputs, and they tested and validated the code before use. The authors take full responsibility for the content of the publication.

R EFERENCES [1] M. A. Cervera, S. R. Soekadar, J. Ushiba, J. d. R. Millán, M. Liu, N. Birbaumer, and G. Garipelli, “Brain-computer interfaces for poststroke motor rehabilitation: a meta-analysis,” Annals of Clinical and Translational Neurology, vol. 5, no. 5, pp. 651–663, 2018. [2] A. Singh, A. A. Hussain, S. Lal, and H. W. Guesgen, “A comprehensive review on critical issues and possible solutions of motor imagery based electroencephalography brain-computer interface,” Sensors, vol. 21, no. 6, p. 2173, 2021. [3] P. Arpaia, D. Coyle, A. Esposito, A. Natalizio, M. Parvis, M. Pesola, and E. Vallefuoco, “Paving the way for motor imagery-based telerehabilitation through a fully wearable BCI system,” Sensors, vol. 23, no. 13, p. 5836, 2023. [4] H. Cho, M. Ahn, S. Ahn, M. Kwon, and S. C. Jun, “EEG datasets for motor imagery brain–computer interface,” GigaScience, vol. 6, no. 7, p. gix034, 2017. [5] D. Wu, Y. Xu, and B.-L. Lu, “Transfer learning for EEG-based brain– computer interfaces: A review of progress made since 2016,” IEEE Transactions on Cognitive and Developmental Systems, vol. 14, no. 1, pp. 4–19, 2020. [6] M. Ahn and S. C. Jun, “Performance variation in motor imagery brain– computer interface: a brief review,” Journal of Neuroscience Methods, vol. 243, pp. 103–110, 2015. [7] A. Myrden and T. Chau, “Effects of user mental state on EEG-BCI performance,” Frontiers in Human Neuroscience, vol. 9, p. 308, 2015. [8] W. Xiong, L. Ma, and H. Li, “Adaptive EEG preprocessing to mitigate electrode shift variability for robust motor imagery classification,” Scientific Reports, vol. 15, no. 1, p. 40808, 2025. [9] M.-H. Lee, O.-Y. Kwon, Y.-J. Kim, H.-K. Kim, Y.-E. Lee, J. Williamson, S. Fazli, and S.-W. Lee, “EEG dataset and OpenBMI toolbox for three BCI paradigms: An investigation into BCI illiteracy,” GigaScience, vol. 8, no. 5, p. giz002, 2019. [10] T. Jia, L. Meng, S. Li, J. Liu, and D. Wu, “Federated motor imagery classification for privacy-preserving brain-computer interfaces,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 32, pp. 3442–3451, 2024. [11] C. Ju, D. Gao, R. Mane, B. Tan, Y. Liu, and C. Guan, “Federated transfer learning for EEG signal classification,” in 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE, 2020, pp. 3040–3045. [12] R. Liu, Y. Chen, A. Li, Y. Ding, H. Yu, and C. Guan, “Aggregating intrinsic information to enhance BCI performance through federated learning,” Neural Networks, vol. 172, p. 106100, 2024. [13] M. Tangermann, K.-R. Müller, A. Aertsen, N. Birbaumer, C. Braun, C. Brunner, R. Leeb, C. Mehring, K. J. Miller, G. R. Müller-Putz, G. Nolte, G. Pfurtscheller, H. Preissl, G. Schalk, A. Schlögl, C. Vidaurre, S. Waldert, and B. Blankertz, “Review of the BCI competition IV,” Frontiers in Neuroscience, vol. 6, p. 55, 2012. [14] K. K. Ang, Z. Y. Chin, C. Wang, C. Guan, and H. Zhang, “Filter bank common spatial pattern algorithm on BCI competition IV datasets 2a and 2b,” Frontiers in Neuroscience, vol. 6, p. 39, 2012. [15] R. T. Schirrmeister, J. T. Springenberg, L. D. J. Fiederer, M. Glasstetter, K. Eggensperger, M. Tangermann, F. Hutter, W. Burgard, and T. Ball, “Deep learning with convolutional neural networks for EEG decoding and visualization,” Human Brain Mapping, vol. 38, no. 11, pp. 5391–5420, 2017. [16] V. J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P. Hung, and B. J. Lance, “EEGNet: A compact convolutional neural network for EEG-based brain–computer interfaces,” Journal of Neural Engineering, vol. 15, no. 5, p. 056013, 2018. [17] R. Mane, E. Chew, K. Chua, K. K. Ang, N. Robinson, A. P. Vinod, S.-W. Lee, and C. Guan, “FBCNet: A multi-view convolutional neural network for brain-computer interface,” arXiv preprint, 2021. [18] X. Jiang, L. Meng, X. Chen, Y. Xu, and D. Wu, “CSP-Net: Common spatial pattern empowered neural networks for EEG-based motor imagery classification,” Knowledge-Based Systems, vol. 305, p. 112668, 2024. [19] H. He and D. Wu, “Transfer learning for brain–computer interfaces: A Euclidean space data alignment approach,” IEEE Transactions on Biomedical Engineering, vol. 67, no. 2, pp. 399–410, 2019. [20] D. Wu, X. Jiang, and R. Peng, “Transfer learning for motor imagery based brain–computer interfaces: A tutorial,” Neural Networks, vol. 153, pp. 235–253, 2022.

[21] D. Wu, “Revisiting Euclidean alignment for transfer learning in EEGbased brain–computer interfaces,” Journal of Neural Engineering, vol. 22, no. 3, p. 031005, 2025. [22] H. Banville, O. Chehab, A. Hyvärinen, D.-A. Engemann, and A. Gramfort, “Uncovering the structure of clinical EEG signals with self-supervised learning,” Journal of Neural Engineering, vol. 18, no. 4, p. 046020, 2021. [23] D. Kostas, S. Aroca-Ouellette, and F. Rudzicz, “BENDR: Using transformers and a contrastive self-supervised learning task to learn from massive amounts of EEG data,” Frontiers in Human Neuroscience, vol. 15, p. 653659, 2021. [24] W.-B. Jiang, L.-M. Zhao, and B.-L. Lu, “Large brain model for learning generic representations with tremendous EEG data in BCI,” in International Conference on Learning Representations, 2024. [25] G. Wang, W. Liu, Y. He, C. Xu, L. Ma, and H. Li, “EEGPT: Pretrained transformer for universal and reliable representation of EEG signals,” Advances in Neural Information Processing Systems, vol. 37, pp. 39 249– 39 280, 2024. [26] J. Wang, S. Zhao, Z. Luo, Y. Zhou, H. Jiang, S. Li, T. Li, and G. Pan, “CBraMod: A criss-cross brain foundation model for EEG decoding,” in International Conference on Learning Representations, 2025. [27] D. Liu, Z. Chen, J. Luo, S. Lian, Y. Chen, S. Hou, X. Zhu, and D. Wu, “MIRepNet: A pipeline and pre-trained model for EEG-based motor imagery classification,” Knowledge-Based Systems, vol. 343, p. 115966, 2026. [28] S. Saeedi, R. Chavarriaga, and J. d. R. Millán, “Long-term stable control of motor-imagery BCI by a locked-in user through adaptive assistance,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 25, no. 4, pp. 380–391, 2016. [29] A. Nagarajan, N. Robinson, K. K. Ang, K. S. G. Chua, E. Chew, and C. Guan, “Transferring a deep learning model from healthy subjects to stroke patients in a motor imagery brain–computer interface,” Journal of Neural Engineering, vol. 21, no. 1, p. 016007, 2024. [30] X. Wang, M. Hersche, B. Tömekce, B. Kaya, M. Magno, and L. Benini, “An accurate EEGNet-based motor-imagery brain–computer interface for low-power edge computing,” in 2020 IEEE International Symposium on Medical Measurements and Applications (MeMeA). IEEE, 2020, pp. 1–6. [31] S. Bian, P. Kang, J. Moosmann, M. Liu, P. Bonazzi, R. Rosipal, and M. Magno, “On-device learning of EEGNet-based network for wearable motor imagery brain-computer interface,” in Proceedings of the 2024 ACM International Symposium on Wearable Computers, 2024, pp. 9–16. [32] W. Hang, J. Li, J. Wang, S. Liang, Q. Wang, B. Lei, G. Li, B. Chen, and J. Qin, “Personalized federated EEG decoding with structured graphbased guidance,” IEEE Transactions on Cognitive and Developmental Systems, vol. 18, no. 4, pp. 1099–1112, 2026. [33] X.-H. Liu, B.-L. Lu, and W.-L. Zheng, “mixEEG: Enhancing EEG federated learning for cross-subject EEG classification with tailored mixup,” in Proceedings of the Annual Meeting of the Cognitive Science Society, 2025. [34] M. G. Arivazhagan, V. Aggarwal, A. K. Singh, and S. Choudhary, “Federated learning with personalization layers,” arXiv preprint, 2019. [35] A. Fallah, A. Mokhtari, and A. Ozdaglar, “Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach,” Advances in Neural Information Processing Systems, vol. 33, pp. 3557– 3568, 2020. [36] L. Collins, H. Hassani, A. Mokhtari, and S. Shakkottai, “Exploiting shared

representations for personalized federated learning,” in International Conference on Machine Learning. PMLR, 2021, pp. 2089–2099. [37] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial Intelligence and Statistics. PMLR, 2017, pp. 1273– 1282. [38] T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V. Smith, “Federated optimization in heterogeneous networks,” Proceedings of Machine Learning and Systems, vol. 2, pp. 429–450, 2020. [39] S. Wang, T. Tuor, T. Salonidis, K. K. Leung, C. Makaya, T. He, and K. Chan, “Adaptive federated learning in resource constrained edge computing systems,” IEEE Journal on Selected Areas in Communications, vol. 37, no. 6, pp. 1205–1221, 2019. [40] F. Lai, X. Zhu, H. V. Madhyastha, and M. Chowdhury, “Oort: Efficient federated learning via guided participant selection,” in 15th USENIX Symposium on Operating Systems Design and Implementation (OSDI 21), 2021, pp. 19–35. [41] C. Xie, S. Koyejo, and I. Gupta, “Asynchronous federated optimization,” arXiv preprint, 2019. [42] J. Nguyen, K. Malik, H. Zhan, A. Yousefpour, M. Rabbat, M. Malek, and D. Huba, “Federated learning with buffered asynchronous aggregation,” in International Conference on Artificial Intelligence and Statistics. PMLR, 2022, pp. 3581–3607. [43] Z. Li, P. Chaturvedi, S. He, H. Chen, G. Singh, V. Kindratenko, E. A. Huerta, K. Kim, and R. Madduri, “FedCompass: Efficient cross-silo federated learning on heterogeneous client devices using a computing power-aware scheduler,” in International Conference on Learning Representations, 2024. [44] X. Li, S. Liu, Z. Zhou, B. Guo, Y. Xu, and Z. Yu, “EchoPFL: Asynchronous personalized federated learning on mobile devices with on-demand staleness control,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 8, no. 1, p. Article 41, 2024. [45] L. Liu, J. Zhang, S. Song, and K. B. Letaief, “Client-edge-cloud hierarchical federated learning,” in ICC 2020–2020 IEEE International Conference on Communications (ICC). IEEE, 2020, pp. 1–6. [46] A. A. Abdellatif, N. Mhaisen, A. Mohamed, A. Erbad, M. Guizani, Z. Dawy, and W. Nasreddine, “Communication-efficient hierarchical federated learning for IoT heterogeneous systems with imbalanced data,” Future Generation Computer Systems, vol. 128, pp. 406–419, 2022. [47] K. Bonawitz, H. Eichner, W. Grieskamp, D. Huba, A. Ingerman, V. Ivanov, C. Kiddon, J. Konečný, S. Mazzocchi, B. McMahan, T. Van Overveldt, D. Petrou, D. Ramage, and J. Roselander, “Towards federated learning at scale: System design,” Proceedings of Machine Learning and Systems, vol. 1, pp. 374–388, 2019. [48] Institute for Knowledge Discovery, Graz University of Technology, “BCI competition IV, dataset 2a: four-class motor imagery,” BCI Competition IV / BNCI Horizon 2020, Dataset 001-2014, 2014, accessed 17 Aug 2026. [Online]. Available: https://bnci-horizon-2020.eu/database/data-sets [49] M.-H. Lee, O.-Y. Kwon, Y.-J. Kim, H.-K. Kim, Y.-E. Lee, J. Williamson, S. Fazli, and S.-W. Lee, “Supporting data for "EEG dataset and OpenBMI toolbox for three BCI paradigms: An investigation into BCI illiteracy",” 2019. [Online]. Available: https://doi.org/10.5524/100542 [50] D. A. Worae and A. Nagarajan, “NEXUS-MI: Communicationaware federated personalization for gateway-coordinated motor-imagery brain-computer interfaces,” Software, version 1.0.0, aug 2026, code repository: https://github.com/WadElla/NEXUS-MI. [Online]. Available: https://doi.org/10.5281/zenodo.22074301

S UPPLEMENTARY M ATERIAL S1. C OMPLETE P OLICY O PERATING P OINTS reach the coordinator and which backbone versions remain Table S1 reports the complete operating-point results for available to the gateways. Consequently, policy changes alter the ideal-link reference and Policies P1–P6 under the default the collaborative training trajectory rather than merely changing heterogeneous-link condition. Accuracy is first averaged equally a communication cost measured after training. P5 and P6 across the three Session 2 calibration budgets, k ∈ {15, 20, 30}, produce the lowest server-to-client traffic in both datasets within each subject and is then averaged across subjects. because they use stale-aware backbone downloading. In the Supplementary Figure S1 complements these budget-averaged primary controlled realization, P5 reduced server-to-client operating points by showing mean held-out decoding accuracy backbone traffic relative to P3 by 41.34% on BCICIV-2a separately at each calibration budget for both datasets and and 41.38% on OpenBMI, corresponding to decreases from learning regimes. The table and figure report the primary 6.27 to 3.68 MB and from 37.49 to 21.98 MB, respectively. The associated accuracy differences vary by dataset and controlled realization using seed 2026. Together, Table S1 and Supplementary Figure S1 characterize regime, which motivates the matched repeated-run analysis each policy as a joint learning and synchronization operating in Section S2. The repeated analysis is restricted to P3 and P5; point. Scheduling, buffering, stale-update admission, and the P1, P2, P4, and P6 operating points therefore correspond backbone-download control determine which local updates to the primary controlled realization.

BCICIV-2a, SB-PH (n=9)

BCICIV-2a, EIB-PH (n=9) 74.0

Accuracy (%)

72.0

72.0

70.0 68.0

70.0

66.0

68.0

64.0

66.0

62.0

64.0

60.0

62.0

Accuracy (%)

OpenBMI, SB-PH (n=54)

OpenBMI, EIB-PH (n=54)

75.0

83.0

74.0

82.0

73.0

81.0

72.0

80.0

71.0

79.0

70.0 78.0 15

20

30

15

Session-2 calibration trials per class

Ideal

P1

20

30

Session-2 calibration trials per class

P2

P3

P4

P5

P6

Fig. S1. Mean held-out decoding accuracy across Session 2 calibration budgets for the ideal-link reference and for Policies P1–P6 under the default heterogeneous-link condition. Each point denotes the cohort mean across subjects (BCICIV-2a, n = 9; OpenBMI, n = 54) from the primary controlled realization using seed 2026. Panels use independently scaled y-axes to show within-panel differences.

TABLE S1 ACCURACY AND COMMUNICATION OPERATING POINTS UNDER THE IDEAL - LINK REFERENCE AND THE DEFAULT HETEROGENEOUS - LINK CONDITION . Dataset

Regime

BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a

SB-PH SB-PH SB-PH SB-PH SB-PH SB-PH SB-PH EIB-PH EIB-PH EIB-PH EIB-PH EIB-PH EIB-PH EIB-PH

OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI

SB-PH SB-PH SB-PH SB-PH SB-PH SB-PH SB-PH EIB-PH EIB-PH EIB-PH EIB-PH EIB-PH EIB-PH EIB-PH

Policy

Subject accuracy (%, mean ± SD)

C2S update traffic (MB)

S2C backbone traffic (MB)

Coordinatorrejected uploads (%)

Mean acceptedupdate staleness (versions)

Mean bufferedupload delay (rounds)

Ideal link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link Ideal link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link

Ideal P1 P2 P3 P4 P5 P6 Ideal P1 P2 P3 P4 P5 P6

67.77 ± 11.67 67.26 ± 11.17 68.26 ± 9.37 68.13 ± 11.39 68.88 ± 10.58 66.77 ± 10.29 64.51 ± 9.02 66.82 ± 13.15 68.24 ± 11.28 67.95 ± 13.09 68.47 ± 10.05 69.80 ± 9.52 68.44 ± 9.61 68.44 ± 10.51

9.30 6.29 9.02 9.02 7.98 8.44 7.40 9.30 6.29 9.02 9.02 7.98 8.44 7.40

9.23 6.27 6.27 6.27 6.27 3.68 3.68 9.23 6.27 6.27 6.27 6.27 3.68 3.68

0.00 0.00 0.00 19.20 8.86 24.59 14.20 0.00 0.00 0.00 19.20 8.86 24.59 14.20

0.00 0.00 0.97 0.27 0.27 0.52 0.52 0.00 0.00 0.97 0.27 0.27 0.52 0.52

— — 1.50 1.49 1.00 1.49 1.00 — — 1.50 1.49 1.00 1.49 1.00

Ideal link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link Ideal link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link Heterogeneous link

Ideal P1 P2 P3 P4 P5 P6 Ideal P1 P2 P3 P4 P5 P6

72.23 ± 12.73 72.59 ± 12.35 72.00 ± 11.57 71.88 ± 12.00 71.55 ± 12.53 73.02 ± 12.18 72.12 ± 11.83 80.75 ± 12.35 80.81 ± 12.56 80.42 ± 12.38 80.13 ± 12.36 80.22 ± 12.81 81.12 ± 12.26 80.65 ± 12.57

55.00 37.65 52.83 52.83 47.00 52.62 46.79 55.00 37.65 52.83 52.83 47.00 52.62 46.79

54.56 37.49 37.49 37.49 37.49 21.98 21.98 54.56 37.49 37.49 37.49 37.49 21.98 21.98

0.00 0.00 0.00 18.51 8.45 22.03 12.37 0.00 0.00 0.00 18.51 8.45 22.03 12.37

0.00 0.00 0.97 0.25 0.25 0.55 0.55 0.00 0.00 0.97 0.25 0.25 0.55 0.55

— — 1.50 1.50 1.00 1.49 1.00 — — 1.50 1.50 1.00 1.49 1.00

Link condition

Notes: For each subject, accuracy is first averaged equally across the three calibration budgets, k ∈ {15, 20, 30}; the reported accuracy is the cohort mean ± sample SD across these subject-level averages. Accuracy and rejected-upload rates are reported in percent, traffic in decimal MB, accepted-update staleness in backbone versions, and buffered-upload delay in communication rounds. C2S and S2C denote client-to-server and server-to-client, respectively; a dash indicates that no buffered-upload event occurred.

S2. ROBUSTNESS ACROSS T RAINING AND G ATEWAY-AVAILABILITY R EALIZATIONS A. Experimental Design and Statistical Analysis

for the repeated accuracy contrast is a crossed hierarchical bootstrap with 10,000 draws. Replicate pairs and subjects were resampled independently with replacement while preserving the within-subject P3/P5 pairing. Across replicates, training initialization and the realized gateway-availability trace were varied, while the P5 priority-scheduling rule remained fixed. The repeated analysis therefore evaluates robustness across these matched stochastic realizations without attributing runto-run differences to any single source.

The principal P3/P5 comparison was repeated under the default heterogeneous-link condition using five matched replicates for each dataset and learning regime. Across the five replicates, model seeds 2026–2030 were paired with availability-trace seeds 12026–12030 and replicate-specific tie-break seeds 22026–22030 for P5. The P5 priority-scheduling rule itself remained fixed across replicates, with these seeds used only B. Accuracy Robustness to resolve otherwise tied gateway priorities; the gatewayTable S2 summarizes the repeated accuracy results. On group seed remained fixed at 2026. Within each replicate BCICIV-2a, the mean P5-minus-P3 effect is +0.80 percentage pair, P3 and P5 used the same initial model state, subject-by- points under SB-PH and +0.88 percentage points under EIBround availability trace, gateway-group assignment, Session 2 PH. The effect is positive in two of five SB-PH replicates and partition, calibration samples, and stochastic state at the start four of five EIB-PH replicates. The corresponding hierarchical of federated training. For EIB-PH, each matched P3/P5 pair 95% confidence intervals are [−1.04, 3.08] and [−1.16, 2.97] also shared the same pooled Session 1 pretrained backbone. percentage points. On OpenBMI, the mean effects are −0.09 The design comprised five matched P3/P5 replicate pairs for percentage points under SB-PH and −0.07 percentage points each dataset–regime combination, yielding 40 completed policy under EIB-PH, with two positive replicates in each regime. runs in total. Their hierarchical intervals are [−0.78, 0.67] and [−0.53, 0.39] For each subject and policy, accuracy was averaged equally percentage points. All four hierarchical intervals include zero. across k ∈ {15, 20, 30}. Replicate-level cohort accuracy was In the primary controlled OpenBMI analysis, the P5-minusthen obtained by averaging across subjects, and the paired P3 effects were +1.14 percentage points under SB-PH and replicate effect was defined as the P5 cohort mean minus the +0.99 percentage points under EIB-PH. Across the five matched corresponding P3 cohort mean. Sample SD, the observed range, replicates, the corresponding effects ranged from −0.52 to and a Student-t 95% confidence interval summarize variation +0.50 percentage points and from −0.27 to +0.11 percentage across the five replicates for both cohort accuracy and the paired points, respectively. Thus, the positive effects observed in the P5-minus-P3 effect. Because the number of replicate pairs is primary analysis were not reproduced consistently across the small, the Student-t interval is descriptive. The primary interval repeated robustness conditions.

TABLE S2 ACCURACY ROBUSTNESS OF P3 AND P5 ACROSS FIVE MATCHED REPLICATES . A. Replicate-level cohort accuracy Dataset

Regime

Policy

n

Mean subject accuracy across budgets (%)

BCICIV-2a BCICIV-2a

SB-PH SB-PH

P3 P5

9 9

68.00 68.80

1.52 0.99

[66.11, 69.88] [67.57, 70.02]

[66.36, 69.83] [67.67, 69.65]

5 5

BCICIV-2a BCICIV-2a

EIB-PH EIB-PH

P3 P5

9 9

68.61 69.50

1.23 1.50

[67.09, 70.13] [67.63, 71.36]

[67.31, 70.63] [67.90, 71.71]

5 5

OpenBMI OpenBMI

SB-PH SB-PH

P3 P5

54 54

72.24 72.15

0.36 0.34

[71.80, 72.69] [71.72, 72.58]

[71.68, 72.55] [71.66, 72.63]

5 5

OpenBMI OpenBMI

EIB-PH EIB-PH

P3 P5

54 54

80.12 80.05

1.06 1.17

[78.80, 81.44] [78.59, 81.51]

[78.52, 81.10] [78.25, 80.96]

5 5

Sample SD across replicates

Student-t 95% CI (pp)

Crossed hierarchical 95% CI (pp)

Observed replicate range (pp)

Replicates with positive P5–P3 effect

Sample SD across replicates

Student-t 95% CI (%)

Observed replicate range (%)

Number of matched replicates

B. Paired P5-versus-P3 accuracy effects Dataset

Regime

n

Mean paired P5–P3 effect (pp)

BCICIV-2a BCICIV-2a

SB-PH EIB-PH

9 9

+0.80 +0.88

2.06 0.95

[−1.76, 3.35] [−0.29, 2.06]

[−1.04, 3.08] [−1.16, 2.97]

[−1.08, 3.14] [−0.28, 1.90]

2/5 4/5

OpenBMI OpenBMI

SB-PH EIB-PH

54 54

−0.09 −0.07

0.40 0.16

[−0.59, 0.41] [−0.26, 0.12]

[−0.78, 0.67] [−0.53, 0.39]

[−0.52, 0.50] [−0.27, 0.11]

2/5 2/5

Notes: Panel A summarizes replicate-level cohort accuracy, and Panel B reports paired P5-minus-P3 accuracy effects in percentage points. The Student-t intervals summarize variation across the five matched replicates. The crossed hierarchical 95% CIs are obtained from resampling matched replicates and subjects while preserving the within-subject P3/P5 pairing.

TABLE S3 S UBJECT- LEVEL P5/P3 ACCURACY EFFECTS ACROSS FIVE MATCHED REPLICATES . Dataset

Regime

n

Median subject-mean P5–P3 effect (pp)

Observed subject-mean range (pp)

Subjects with positive mean effect

Full subject-by-replicate range (pp)

≤ −5 pp effects

≥ +5 pp effects

BCICIV-2a BCICIV-2a

SB-PH EIB-PH

9 9

+0.83 +0.88

[−0.93, 3.33] [−2.96, 6.02]

6/9 6/9

[−6.94, 10.19] [−7.64, 9.03]

4/45 5/45

6/45 9/45

OpenBMI OpenBMI

SB-PH EIB-PH

54 54

−0.08 −0.22

[−4.22, 4.11] [−3.78, 3.17]

25/54 21/54

[−12.50, 21.94] [−10.56, 7.50]

29/270 9/270

25/270 11/270

Notes: Subject-mean effects average each subject’s paired P5-minus-P3 accuracy effect across the five replicates; positive-subject counts include subjects with a strictly positive mean effect. The full subject-by-replicate range and the ±5-pp counts are calculated over all matched subject-by-replicate effects, yielding 45 comparisons for BCICIV-2a and 270 for OpenBMI. Effects exactly equal to −5 or +5 pp are included in the corresponding threshold counts.

C. Subject-Level Heterogeneity Table S3 summarizes subject-level P5-minus-P3 effects across the five replicates. Positive and negative effects occur in every dataset–regime condition. On BCICIV-2a, the subjectmean effects range from −0.93 to +3.33 percentage points under SB-PH and from −2.96 to +6.02 percentage points under EIB-PH. The corresponding OpenBMI ranges are −4.22 to +4.11 percentage points and −3.78 to +3.17 percentage points. The full subject-by-replicate ranges are wider, reaching [−6.94, 10.19] percentage points for BCICIV-2a SB-PH and [−12.50, 21.94] percentage points for OpenBMI SB-PH. Subject-by-replicate effects of at least five percentage points in both the positive and negative directions were observed under SB-PH and EIB-PH on both datasets. Similarity in cohort means therefore does not imply uniform subject-level behavior across synchronization policies. D. Communication Robustness Table S4 reports communication outcomes across the five matched replicates. The reported communication values are common to SB-PH and EIB-PH at the displayed precision. P5 reduces server-to-client backbone traffic in every replicate. The mean reduction is 42.23% on BCICIV-2a, with an observed range of 41.59% to 42.76%, and 41.98% on OpenBMI, with a range of 41.47% to 42.47%.

P5 also reduces total communication in all five replicates. Relative to P3, mean total traffic decreases by 3.20 MB on BCICIV-2a and 15.91 MB on OpenBMI. These reductions are accompanied by increases in the coordinator-rejected upload rate of 5.29 and 3.91 percentage points, respectively, and by increases in mean accepted-update staleness of 0.266 and 0.299 backbone versions. Mean buffered-upload delay differs by no more than 0.003 rounds on either dataset. The communication savings therefore arise primarily from selective synchronization and avoided backbone downloads rather than from longer buffering intervals. E. Replicate-Level Results and Interpretation Table S5 reports the complete replicate-level cohort accuracies and server-to-client reductions. On BCICIV-2a, the accuracy effect ranges from −1.08 to +3.14 percentage points under SB-PH and from −0.28 to +1.90 percentage points under EIB-PH. On OpenBMI, the corresponding ranges are −0.52 to +0.50 percentage points and −0.27 to +0.11 percentage points. By contrast, the server-to-client reduction remains close to 42% in every replicate. Across the five matched replicates, P5 consistently reduced server-to-client backbone traffic and total communication. The accuracy differences were comparatively small, varied in direction across replicates, and had crossed hierarchicalbootstrap 95% confidence intervals that included zero in all

TABLE S4 C OMMUNICATION ROBUSTNESS OF P3 AND P5 ACROSS FIVE MATCHED REPLICATES . Unit

P3 mean ± SD

P5 mean ± SD

Paired difference mean ± SD

Student-t 95% CI

Observed range

Direction count (+/ − / =)

C2S traffic S2C traffic Total traffic Coordinator-rejected uploads Accepted-update staleness Buffered-upload delay Avoided S2C downloads S2C traffic reduction

MB MB MB pp versions rounds pp %

8.765 ± 0.082 6.279 ± 0.055 15.044 ± 0.066 17.962 ± 1.050 0.244 ± 0.010 1.489 ± 0.018 0.000 ± 0.000 —

8.215 ± 0.101 3.628 ± 0.017 11.843 ± 0.107 23.251 ± 0.956 0.510 ± 0.008 1.487 ± 0.019 37.326 ± 0.424 —

−0.550 ± 0.043 −2.652 ± 0.054 −3.201 ± 0.091 +5.290 ± 0.294 +0.266 ± 0.007 −0.002 ± 0.003 +37.326 ± 0.424 +42.226 ± 0.515

[−0.603, −0.496] [−2.718, −2.585] [−3.314, −3.088] [4.925, 5.655] [0.257, 0.274] [−0.006, 0.002] [36.799, 37.852] [41.587, 42.865]

[−0.603, −0.482] [−2.710, −2.599] [−3.313, −3.081] [4.817, 5.579] [0.255, 0.272] [−0.006, 0.000] [36.667, 37.744] [41.589, 42.762]

0/5/0 0/5/0 0/5/0 5/0/0 5/0/0 0/2/3 5/0/0 5/0/0

C2S traffic S2C traffic Total traffic Coordinator-rejected uploads Accepted-update staleness Buffered-upload delay Avoided S2C downloads S2C traffic reduction

MB MB MB pp versions rounds pp %

52.211 ± 0.195 37.392 ± 0.417 89.604 ± 0.606 17.810 ± 0.524 0.249 ± 0.006 1.489 ± 0.012 0.000 ± 0.000 —

51.994 ± 0.153 21.696 ± 0.148 73.690 ± 0.290 21.716 ± 0.406 0.548 ± 0.004 1.487 ± 0.011 42.154 ± 0.345 —

−0.217 ± 0.050 −15.697 ± 0.303 −15.914 ± 0.345 +3.906 ± 0.282 +0.299 ± 0.004 −0.003 ± 0.002 +42.154 ± 0.345 +41.976 ± 0.380

[−0.280, −0.155] [−16.074, −15.320] [−16.343, −15.486] [3.557, 4.256] [0.294, 0.303] [−0.005, −0.001] [41.725, 42.583] [41.505, 42.447]

[−0.272, −0.136] [−16.083, −15.401] [−16.354, −15.536] [3.603, 4.275] [0.294, 0.304] [−0.005, 0.000] [41.644, 42.566] [41.466, 42.465]

0/5/0 0/5/0 0/5/0 5/0/0 5/0/0 0/4/1 5/0/0 5/0/0

Dataset

Metric

BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI

Notes: P3 and P5 columns report replicate-level mean ± sample SD. Paired differences are computed as P5 minus P3, except S2C traffic reduction, which is measured relative to P3 and is positive when P5 uses less S2C traffic. Direction counts (+/ − / =) give the number of matched replicates with positive, negative, or zero P5-minus-P3 differences; for S2C traffic reduction, a positive direction denotes lower S2C traffic under P5.

TABLE S5 C OMPLETE REPLICATE - LEVEL P3/P5 RESULTS ACROSS FIVE MATCHED REPLICATES . Dataset

Regime

Model seed

Availabilitytrace seed

P5 scheduler tie-break seed

P3 mean subject accuracy (%)

P5 mean subject accuracy (%)

P5–P3 accuracy difference (pp)

S2C traffic reduction (%)

BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a BCICIV-2a

SB-PH SB-PH SB-PH SB-PH SB-PH EIB-PH EIB-PH EIB-PH EIB-PH EIB-PH

2026 2027 2028 2029 2030 2026 2027 2028 2029 2030

12026 12027 12028 12029 12030 12026 12027 12028 12029 12030

22026 22027 22028 22029 22030 22026 22027 22028 22029 22030

66.36 68.57 69.83 68.75 66.49 68.36 68.57 70.63 67.31 68.18

69.50 67.77 69.65 67.67 69.39 70.27 68.67 71.71 68.93 67.90

+3.14 −0.80 −0.18 −1.08 +2.91 +1.90 +0.10 +1.08 +1.62 −0.28

41.82 42.76 42.29 42.67 41.59 41.82 42.76 42.29 42.67 41.59

OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI OpenBMI

SB-PH SB-PH SB-PH SB-PH SB-PH EIB-PH EIB-PH EIB-PH EIB-PH EIB-PH

2026 2027 2028 2029 2030 2026 2027 2028 2029 2030

12026 12027 12028 12029 12030 12026 12027 12028 12029 12030

22026 22027 22028 22029 22030 22026 22027 22028 22029 22030

72.55 71.68 72.55 72.27 72.18 80.81 79.58 80.61 81.10 78.52

72.63 72.18 72.17 72.12 71.66 80.86 79.47 80.72 80.96 78.25

+0.09 +0.50 −0.38 −0.15 −0.52 +0.06 −0.11 +0.11 −0.14 −0.27

42.47 41.81 42.20 41.47 41.95 42.47 41.81 42.20 41.47 41.95

Notes: Cohort accuracy first averages each subject equally across the three calibration budgets and then averages across subjects. Accuracy differences are computed as P5 minus P3, while S2C traffic reductions are measured relative to P3. The gateway-group seed is fixed at 2026 across all replicates.

four dataset–regime conditions. Taken together, these results show that the communication reduction is consistent across the evaluated replicates, whereas the accuracy effect remains realization-dependent. S3. L INK -AVAILABILITY S ENSITIVITY Table S6 reports the EIB-PH results for P3 and P5 under mild, default, and severe heterogeneous-link profiles. The online probabilities assigned to the high-, moderate-, and low-availability gateway groups are 0.98/0.85/0.60, 0.95/0.70/0.40, and 0.90/0.50/0.20, respectively. P3 and P5 retain the same FIFO buffering and stale-update admission rules in all three profiles. Their comparison therefore continues to isolate the combined effect of communication-aware scheduling and stale-aware backbone downloading. As availability decreases, both policies encounter fewer synchronization opportunities and a larger fraction of transmitted uploads is rejected by the coordinator. From the mild to the severe profile, the rejected-upload rate rises from 9.19% to

28.25% under P3 and from 15.39% to 32.20% under P5 on BCICIV-2a. The corresponding OpenBMI rates rise from 8.63% to 27.52% and from 13.60% to 30.64%. Mean buffered-upload delay reaches approximately 1.6 rounds in the severe profile, and mean accepted-update staleness remains higher under P5. Despite these changes, P5 uses less server-to-client traffic at every severity level. Its traffic is 4.07, 3.68, and 2.99 MB on BCICIV-2a, compared with 7.46, 6.27, and 4.86 MB under P3. On OpenBMI, the corresponding P5 values are 24.39, 21.98, and 17.81 MB, compared with 44.36, 37.49, and 28.99 MB under P3. The accuracy contrast does not follow a common direction across profiles. On BCICIV-2a, the P5-minus-P3 differences are +1.77, −0.03, and +0.16 percentage points under mild, default, and severe availability. On OpenBMI, the corresponding differences are +0.08, +0.99, and −0.46 percentage points. The severity analysis therefore shows a persistent communication reduction but no profile-invariant accuracy advantage. This sensitivity analysis should be distinguished from the repeated

TABLE S6 S ENSITIVITY OF EIB-PH TO HETEROGENEOUS GATEWAY AVAILABILITY. Availability profile (high/moderate/low)

Dataset

Policy

Mean subject accuracy (%)

C2S update traffic (MB)

S2C backbone traffic (MB)

Coordinator-rejected uploads (%)

Mean acceptedupdate staleness (versions)

Mean bufferedupload delay (rounds)

BCICIV-2a BCICIV-2a

Mild (0.98/0.85/0.60) Mild (0.98/0.85/0.60)

P3 P5

68.29 70.06

9.18 7.78

7.46 4.07

9.19 15.39

0.20 0.45

1.35 1.35

BCICIV-2a BCICIV-2a

Default (0.95/0.70/0.40) Default (0.95/0.70/0.40)

P3 P5

68.47 68.44

9.02 8.44

6.27 3.68

19.20 24.59

0.27 0.52

1.49 1.49

BCICIV-2a BCICIV-2a

Severe (0.90/0.50/0.20) Severe (0.90/0.50/0.20)

P3 P5

67.90 68.06

7.96 7.84

4.86 2.99

28.25 32.20

0.29 0.54

1.62 1.60

OpenBMI OpenBMI

Mild (0.98/0.85/0.60) Mild (0.98/0.85/0.60)

P3 P5

80.53 80.61

54.66 50.73

44.36 24.39

8.63 13.60

0.21 0.51

1.33 1.33

OpenBMI OpenBMI

Default (0.95/0.70/0.40) Default (0.95/0.70/0.40)

P3 P5

80.13 81.12

52.83 52.62

37.49 21.98

18.51 22.03

0.25 0.55

1.50 1.49

OpenBMI OpenBMI

Severe (0.90/0.50/0.20) Severe (0.90/0.50/0.20)

P3 P5

81.19 80.73

45.95 45.95

28.99 17.81

27.52 30.64

0.26 0.53

1.60 1.59

Notes: Availability profiles report the high/moderate/low gateway online probabilities. Mean accepted-update staleness is averaged across accepted uploads, while mean buffered-upload delay is averaged across buffered uploads that reach stale-update admission.

analysis in Section S2: here, availability severity is varied within the primary controlled realization, whereas the repeated analysis evaluates five matched P3/P5 replicates under the default availability profile with different training initializations and realized gateway-availability traces, while retaining the same P5 priority-scheduling rule with replicate-specific seeded tie resolution.

with the subject-level accuracy difference under both P3 and P5. However, the corresponding 95% bootstrap confidence intervals are wide and include zero: [−1.00, 0.11] for P3 and [−1.00, 0.07] for P5. The remaining BCICIV-2a associations are smaller in magnitude and likewise have confidence intervals that include zero. On OpenBMI, all point estimates are small, ranging from −0.17 to 0.09, and all corresponding confidence intervals also include zero. The avoided-download correlation S4. A SSOCIATIONS W ITH S UBJECT-L EVEL ACCURACY is undefined for P3 because this quantity does not vary across D IFFERENCES subjects under the P3 download rule. Table S7 reports descriptive Spearman rank correlations Across the evaluated datasets and policies, no recorded between each subject’s EIB-PH accuracy difference from the communication-exposure quantity shows a consistent associmatched ideal-link reference and six candidate quantities: ideal- ation with subject-level accuracy differences. In particular, link accuracy, assigned gateway online probability, coordinator- the present correlations do not support attributing subjectrejected uploads, mean accepted-update staleness, updates level degradation to rejected uploads, accepted-update stalplaced in the buffer, and avoided backbone downloads. Uncer- eness, buffering, gateway availability, or avoided downloads tainty is summarized using 95% percentile confidence intervals individually. The results should therefore be interpreted as from 10,000 subject-level bootstrap resamples within each exploratory descriptions of subject-level covariation rather than dataset–policy analysis. These analyses characterize monotonic as evidence of causal mechanisms or reliable predictors of associations within the primary controlled realization and do vulnerability. This caution is especially important for BCICIVnot establish causal relationships or isolate the contribution of 2a because the analysis includes only nine subjects, and the any individual communication quantity. reported confidence intervals are not adjusted for the multiple The largest point estimates occur on BCICIV-2a, where correlations examined. ideal-link accuracy has a Spearman correlation of ρ = −0.68 TABLE S7 A SSOCIATIONS OF BASELINE ACCURACY AND COMMUNICATION EXPOSURE WITH SUBJECT- LEVEL ACCURACY DIFFERENCES UNDER EIB-PH. Spearman correlation with subject-level accuracy difference (ρ; 95% CI) Dataset

Policy

n

BCICIV-2a

P3

9

BCICIV-2a

P5

9

OpenBMI

P3

54

OpenBMI

P5

54

Ideal-link accuracy

Assigned gateway online probability

Coordinator-rejected uploads (count)

Mean accepted-update staleness (versions)

Updates placed in buffer (count)

Avoided backbone downloads (count)

−0.68 [−1.00, 0.11] −0.68 [−1.00, 0.07]

0.26 [−0.57, 0.83] 0.21 [−0.59, 0.82]

−0.08 [−0.74, 0.74] 0.03 [−0.72, 0.73]

−0.25 [−0.83, 0.64] 0.13 [−0.71, 0.72]

−0.14 [−0.83, 0.74] 0.00 [−0.75, 0.71]

0.05 [−0.64, 0.83]

−0.17 [−0.42, 0.10] −0.14 [−0.42, 0.16]

0.09 [−0.19, 0.37] 0.05 [−0.23, 0.31]

−0.07 [−0.35, 0.21] 0.03 [−0.24, 0.32]

−0.10 [−0.38, 0.19] −0.11 [−0.36, 0.16]

−0.08 [−0.36, 0.20] −0.02 [−0.28, 0.24]

0.03 [−0.23, 0.29]

—

—

Notes: Entries report Spearman rank correlations with each policy’s subject-level accuracy difference from the matched ideal-link reference, with 95% percentile CIs obtained from 10,000 subject-level bootstrap resamples (seed 2026) within each dataset–policy analysis. The CIs are not adjusted for multiple comparisons, and the associations are descriptive rather than causal. A dash indicates that the correlation is undefined because the corresponding quantity does not vary across subjects.