arXiv CSopen access

CogAdapt: Transferring Clinical ECG Foundation Models to Wearable Cognitive Load Assessment via Lead Adaptation

arXiv:2605.22774v1 [cs.LG] 21 May 2026

CogAdapt: Transferring Clinical ECG Foundation Models to Wearable Cognitive Load Assessment via Lead Adaptation Amir Mousavi1 , Mohammad Sadegh Sirjani1 , Erfan Nourbakhsh1 , Mimi Xie1 and Rocky Slavin1 , Leslie Neely3 , John Davis2 , John Quarles1 1 Department of Computer Science, College of AI, Cyber and Computing, The University of Texas at San Antonio 2 Department of Educational Psychology, College of Education and Human Development, The University of Texas at San Antonio 3 Department of Neuroscience, Developmental and Regenerative Biology, College of Sciences, The University of Texas at San Antonio {seyedamir.mousavi, mohammadsadegh.sirjani, erfan.nourbakhsh, mimi.xie, rocky.slavin, john.quarles}@utsa.edu, {leslie.neely, john.davis2}@utsa.edu Abstract Real-time cognitive load assessment is essential for adaptive human-computer interaction but remains challenging due to limited labeled data and poor cross-subject generalization. Recent ECG foundation models pre-trained on millions of clinical recordings offer rich representations, but cannot be directly applied to wearable devices due to sensor configuration mismatch and task differences. In this paper, we propose CogAdapt, a framework that adapts clinical ECG foundation models to wearable cognitive load assessment. CogAdapt introduces LeadBridge, a learnable adapter that transforms 3lead wearable signals into anatomically-consistent 12-lead representations, and ProFine, a progressive fine-tuning strategy that gradually unfreezes encoder layers while preventing catastrophic forgetting. Evaluations on two public datasets (CLARE and CL-Drive) under leave-one-subject-out crossvalidation show that CogAdapt substantially outperforms baselines trained from scratch, achieving macro-F1 scores of 0.626 and 0.768. These results demonstrate the promise of foundation model adaptation for subject-independent cognitive load assessment from wearable sensors.

1

Introduction

Real-time cognitive load assessment enables adaptive humancomputer interaction systems in education, training, and driving contexts [Kundu, 2022; Kosch et al., 2023]. However, three fundamental challenges hinder adoption [Gkintoni et al., 2025; Suryani et al., 2024]. First, retrospective labeling introduces memory biases, as traditional approaches collect subjective ratings only after task completion [Naismith and Cavalcanti, 2015], producing labels reflecting aggregate impressions rather than momentary cognitive states. Second, sparse temporal annotations limit real-time model develop-

Figure 1: The core challenge: Pre-trained foundation models expect rich, clinical 12-lead data, but practical applications rely on sparse, noisy signals from wearable 3-lead devices

ment, with most datasets providing only one label per multiminute trial. Third, limited training data constrains generalization, as datasets typically contain fewer than 25 participants with only a few hours of data each. Recent datasets have begun addressing these limitations. CLARE [Bhatti et al., 2024] captures multi-modal physiological signals from 24 participants with self-reported load levels every 10 seconds, while CL-Drive [Angkan et al., 2024] records ECG and other sensors from 21 drivers in a simulator. Despite providing fine-grained labeling, the limited participant counts and recording hours present challenges for training deep neural networks without overfitting. Domain shift between controlled studies and real-world settings further necessitates subject-independent modeling. We therefore focus on leave-one-subject-out cross-validation, which reveals substantial performance degradation compared to within-subject testing—for instance, a CNN on CLARE achieves 78.4% accuracy under k-fold validation but drops to 63.3% under leave-one-subject-out evaluation [Bhatti et al., 2024].

Electrocardiography (ECG) provides a suitable biosignal for cognitive load assessment due to its non-invasive nature and sensitivity to autonomic nervous system changes during mental effort. Clinical ECG systems use 12 leads to capture comprehensive cardiac activity, while wearable devices employ simplified 3-lead configurations, creating a gap when leveraging clinical ECG models for real-world applications. We address these challenges by adapting a pre-trained ECG foundation model to cognitive load classification. Foundation models trained via self-supervised learning on large datasets have transformed natural language processing, computer vision, and biosignal analysis [Gu et al., 2025; Han et al., 2024]. ECG-FM is a transformer-based foundation model trained on clinical ECG recordings that achieves strong performance on cardiac diagnosis tasks [McKeen et al., 2025]. We hypothesize that ECG signatures of cognitive load, such as heart rate variability changes from sympathetic nervous system activity [Bhatti et al., 2024], can be detected more effectively by transferring knowledge from ECG-FM than by training task-specific models from scratch. This represents a non-trivial transfer: adapting models trained to detect cardiac pathologies in clinical settings to assess mental states from consumer-grade wearable sensors. Applying ECG-FM to wearable cognitive load assessment requires addressing two domain gaps (Figure 1): (i) Sensor configuration mismatch—ECG-FM expects 12-lead clinical ECGs while CLARE provides only 3-lead wearable data; and (ii) Task mismatch—the pre-training objective of cardiac abnormality detection differs from cognitive state estimation. Since cognitive load datasets lack 12-lead recordings for supervision, we leverage the PTB-XL dataset [Wagner et al., 2020], containing 21,837 clinical 12-lead ECG recordings, to pre-train our adapter on normal sinus rhythm recordings before freezing it during downstream training. We propose CogAdapt, a framework for transferring clinical ECG foundation models to wearable cognitive load classification. Our main contributions are: • We propose LeadBridge, a lightweight learnable adapter that employs Wilson Central Terminal rereferencing to transform 3-lead wearable inputs into anatomically-consistent 12-lead representations compatible with clinical foundation models. • We develop ProFine, a progressive fine-tuning strategy that gradually unfreezes encoder layers with tier-wise learning rate decay, balancing task adaptation against catastrophic forgetting. • Experiments on two cognitive load datasets under leaveone-subject-out evaluation demonstrate that CogAdapt outperforms baselines trained from scratch, establishing foundation model adaptation as a viable approach for wearable cognitive state assessment.

2

Related Work

Cognitive Load Datasets and Ground Truth Early cognitive load datasets [Gjoreski et al., 2020; Markova and others, 2019] provided only retrospective NASA-TLX ratings or block-level labels, lacking the temporal granularity required

for real-time monitoring. Recent datasets address these limitations through dense temporal supervision. CLARE captures multimodal signals from 24 participants during MATBII multitasking with in-situ ratings every 10 seconds, while CL-Drive records data from 21 drivers with 10-second workload ratings during simulator scenarios. Despite fine-grained labeling, limited participant counts and substantial intersubject variability degrade leave-one-subject-out (LOSO) performance. A comprehensive dataset comparison is provided in Appendix B. Self-Supervised ECG Foundation Models Selfsupervised learning (SSL) enables the development of general-purpose ECG models [Gu et al., 2025; Han et al., 2024]. SSL approaches utilize contrastive frameworks [Chen et al., 2020] or exploit inherent signal structure through methods like CLOCS [Kiyasseh et al., 2021] and SimPer [Yang et al., 2022]. ECG-FM, a 91M-parameter transformer pre-trained on 1.5M clinical 12-lead ECGs, produces 768-dimensional embeddings with expert-level diagnostic accuracy. SSL-based models may transfer better to cognitive load tasks than supervised arrhythmia models, as they capture broader morphological patterns without task-specific bias [Yang et al., 2023a; Abbaspourazad et al., 2024], potentially preserving subtle physiological variations like sympathetic-driven heart rate variability [Solhjoo et al., 2019]. However, adapting clinical foundation models to wearable devices requires addressing substantial domain gaps, including sensor configuration differences and signal quality variations. Bridging Clinical and Wearable ECG Domains Clinical 12-lead systems use standardized electrode placements while wearables employ simplified 3-lead configurations with variable positioning across different devices. Traditional fixed transforms [Vozda and Cerny, 2015] fail with non-standard placements [Feild et al., 2008], while learned neural mappings outperform static approaches by adapting to placement variability [Hassannia and Sameni, 2025]. Wearable recordings exhibit lower signal-to-noise ratios from motion artifacts and environmental interference [Satija et al., 2018], requiring noise augmentation [Rahman et al., 2022], deep denoising [Mvuh et al., 2024], and domain alignment techniques [Shang et al., 2021]. Window-wise normalization prevents data leakage during LOSO evaluation by computing statistics independently for each segment [Yang et al., 2023b]. CogAdapt addresses these challenges through LeadBridge, combining learnable lead transformation with Wilson Central Terminal re-referencing and window-wise normalization to bridge the clinical-wearable domain gap. ECG-Based Cognitive Load Classification Cardiovascular dynamics reliably indicate cognitive load through autonomic nervous system modulation [Hughes et al., 2019; Ayres et al., 2021]. Traditional approaches extract handcrafted HRV features [Raza et al., 2024], achieving 80% LOSO accuracy but requiring extensive feature engineering and domain expertise [Kalatzis and others, 2021]. Deep learning methods [Kim et al., 2022] learn representations end-to-end without manual feature design, with ECG emerging as the strongest single modality on CLARE and CL-

Figure 2: CogAdapt architecture bridging wearable 3-lead ECG to clinical 12-lead foundation models. The framework consists of data processing (WCT re-referencing and normalization), LeadBridge adapter (learnable 3 → 12 lead transformation), pre-trained ECG-FM encoder, and classification head with focal loss for cognitive load assessment.

Drive datasets. However, limited data (fewer than 25 subjects) risks overfitting to dataset-specific artifacts rather than generalizable cognitive load patterns. Transfer learning addresses data scarcity through pre-training on large-scale ECG datasets [Han et al., 2025], with multitask learning demonstrating advantages over single-task training [Gjoreski et al., 2020]. CogAdapt advances this approach by adapting the ECG-FM foundation model via LeadBridge and ProFine, leveraging 1.5M pre-training samples to enable robust subject-independent evaluation.

3

Methodology

3.1

Framework Overview

Figure 2 presents CogAdapt, a transfer learning framework that adapts a clinical ECG foundation model to wearable cognitive load assessment. CogAdapt addresses two key gaps between clinical and wearable settings: (i) the mismatch between 3-lead wearable recordings and the 12-lead input expected by clinical foundation encoders, and (ii) the task shift from cardiac diagnosis to cognitive state estimation. The framework consists of four sequential components: • Data Processing Pipeline (Section 3.3): preprocessing, Wilson Central Terminal (WCT) re-referencing, windowing, and normalization; • LeadBridge Adapter (Section 3.4): a learnable 3 → 12 lead transformation that aligns wearable signals with clinical input format; • ProFine Fine-Tuning Strategy (Section 3.6): progressive adaptation via three training scenarios with different parameter-efficiency trade-offs;

• Classification Head and Objective (Section 3.7): a lightweight classifier trained with focal loss and automatic class rebalancing. Together, these modules enable CogAdapt to preserve the strengths of large-scale clinical pre-training while remaining compatible with wearable sensing constraints.

3.2

Problem Formulation

We formulate cognitive load assessment as supervised timeseries classification under heterogeneous domain shift. Let X ∈ RC×T denote an ECG segment with C channels and length T , and let Y = {0, 1} denote binary labels corresponding to low vs. high cognitive load. In the source domain DS , a foundation model encoder Eθ is pre-trained on clinical 12-lead ECGs (CS = 12). In the target domain DT = {(xi , yi )}N i=1 , only wearable 3-lead ECGs are available (CT = 3), i.e., xi ∈ R3×T . Transfer challenges. Direct transfer is hindered by two fundamental obstacles: 1. Architecture mismatch. The encoder expects 12 channels, whereas wearable devices provide only three. Naive lead expansion (e.g., zero-padding) introduces missing-channel artifacts and fails to preserve inter-lead structure learned in clinical pre-training. 2. Representation drift. Clinical pre-training optimizes representations for pathology-related cues, while cognitive load manifests as subtler autonomic modulations. Effective transfer therefore requires task-aligned feature adaptation.

Learning objective. Our goal is to learn a mapping Φ : R3×T → Y with parameters Θ = {θ, ϕ, ψ}, where ϕ parameterizes the LeadBridge adapter Aϕ and ψ parameterizes the classification head gψ . CogAdapt resolves the architecture mismatch via LeadBridge and mitigates representation drift via progressive fine-tuning (ProFine).

3.3

Wearable ECG Data Processing

As shown in the top panel of Figure 2, CogAdapt first converts raw wearable ECG streams into fixed-length windows suitable for downstream feature extraction. Signal Preprocessing We resample all raw ECG signals to a common sampling rate fs , followed by bandpass filtering to suppress baseline wander and high-frequency noise. Samples missing all leads are discarded. For sporadic missing values, we apply zero-filling. If a lead becomes constant within a window (flat-line behavior), it is set to zero to prevent degenerate inputs. WCT-Based Re-referencing The wearable configuration provides Lead I (LA−RA), Lead II (LL − RA), and a chest lead VxRL referenced to the right leg. Following the CLARE convention [Bhatti et al., 2024], we express the chest lead relative to Wilson’s Central Terminal (WCT), defined as the average of limb electrode potentials and assumed to remain approximately steady during the cardiac cycle [Moeinzadeh et al., 2018]. In standard acquisition, the right leg electrode functions as a driven ground, and the cardiac component of the RL and RA potentials can be treated as approximately equivalent [Kligfield et al., 2007]. Under RL ≈ RA, we apply WCT correction and obtain the re-referenced 3-lead input: xref = [I, II, VxWCT ]⊤ Full derivations and the reconstruction procedure are provided in Appendix C.1. Temporal Segmentation and Label Assignment We segment continuous recordings into fixed-length windows xk ∈ R3×T of duration w seconds. Cognitive load labels are provided at ∆-second intervals (e.g., every 10s). For each window Wk = [tk , tk +w), we assign the label corresponding to the interval that contains the window midpoint: j t + w/2 k k y(Wk ) = yi⋆ , i⋆ = (1) ∆ This midpoint rule avoids ambiguous assignments at interval boundaries [Lawhern et al., 2013; Potharaju et al., 2025]. During training, we use overlapping windows (stride s < w) to reduce position-dependent overfitting. During evaluation, we use non-overlapping windows aligned with label intervals for consistent reporting. Window-Level Normalization Each window is normalized per lead using z-score normalization: xk − µ k x̃k = (2) σk + ϵ where µk , σ k ∈ R3 are per-window per-lead statistics and ϵ is a small constant.

Figure 3: Progressive fine-tuning scenarios. Scenario A (left) freezes the entire encoder; Scenario B (middle) unfreezes top layers; Scenario C (right) unfreezes all parameters.

3.4

LeadBridge: Learnable 3 → 12 Lead Adaptation

To align 3-lead wearable inputs with the 12-lead interface expected by the clinical foundation encoder, we propose LeadBridge, a lightweight adapter mapping R3×T → R12×T (Figure 2). Because cognitive load datasets (CLARE, CL-Drive) provide only wearable 3-lead recordings and do not include 12lead ground truth, the adapter cannot be directly supervised during downstream cognitive load training. We therefore pretrain LeadBridge on PTB-XL, which contains 21,837 clinical 12-lead ECG recordings sampled at 500 Hz. PTB-XL Pre-training Setup We filter PTB-XL to retain normal sinus rhythm recordings (SCP code NORM ≥ 80% confidence), yielding approximately 9,500 clean examples. Each 10-second recording is segmented into 5,000-sample windows (10s at 500 Hz) with 50% overlap. We randomly split the data into 80%/20% train/validation subsets. After pre-training, LeadBridge is integrated into CogAdapt and further fine-tuned end-to-end with classification loss during downstream training. This task-specific adaptation allows the adapter to learn 12-lead representations optimized for cognitive load states rather than reconstruction fidelity alone, while PTB-XL pre-training provides anatomically-informed initialization that accelerates convergence and improves generalization. Adapter Architecture LeadBridge implements a multi-layer convolutional network that progressively transforms 3-lead inputs into 12-lead representations: Aϕ (x) := W3 · σ(BN(W2 · σ(BN(W1 x))))

(3)

where W1 ∈ R64×3 , W2 ∈ R64×64 , W3 ∈ R12×64 , σ(·) denotes ReLU, and BN(·) denotes batch normalization. In practice, each projection is implemented using a Conv1D layer (kernel size 1), followed by BatchNorm1D and ReLU for the first two layers, with the final layer producing the 12-lead output. While standard clinical leads satisfy deterministic relationships (e.g., Einthoven’s law), wearable electrode placements can deviate from clinical standards. Instead of enforcing fixed lead equations, LeadBridge learns a data-driven mapping from PTB-XL supervision, improving robustness to

placement variability while aligning signal statistics to the encoder’s pre-training distribution. Initialization and normalization. We initialize W using Kaiming initialization [He et al., 2015]. Batch normalization [Ioffe, 2015] further reduces covariate shift between adapted wearable signals and clinical pre-training inputs.

3.5

Forward Inference Through CogAdapt

Given a normalized window xk ∈ R3×T , CogAdapt computes predictions as: x12,k = Aϕ (xk )

(4)

Hk = Eθ (x12,k )

(5)

zk = Pool(Hk ) ∈ Rd

(6) 2

pk = softmax(gψ (zk )) ∈ R

(7)

Here, Eθ is the pre-trained ECG-FM transformer encoder, Pool(·) denotes temporal mean pooling, and gψ is the classification head. Since all windows have fixed duration, padding masks are not required.

3.6

ProFine: Progressive Fine-Tuning Strategy

CogAdapt adapts the foundation encoder to cognitive load classification using ProFine, a progressive fine-tuning strategy designed to balance adaptation capacity against catastrophic forgetting. ProFine defines three scenarios with increasing trainable parameters (Figure 3). Let θ denote encoder parameters decomposed into L layers {ℓ1 , . . . , ℓL }, where ℓL is nearest to the output. Scenario A: Frozen Encoder (Linear Probing) Scenario A freezes the entire encoder and trains only the downstream modules: ΘA train = {ϕ, ψ}

(8)

This setting evaluates whether fixed foundation representations are sufficient for cognitive load assessment when only lightweight adaptation is permitted. Since the trainable modules are randomly initialized, we apply data augmentation for regularization. Scenario B: Partial Unfreeze (Top Encoder Layers) Scenario B unfreezes only the top K layers: ΘB train = {ϕ, ψ} ∪ {ℓj | j > L − K}

(9)

With K=2, this allows joint adaptation of the LeadBridge mapping, upper encoder layers, and classification head are trained together, and allows task-specific adaptation while keeping low-level ECG morphology intact. Scenario C: Full Fine-Tuning (All Encoder Layers) Scenario C performs end-to-end optimization: ΘC train = Θ = {ϕ, ψ, ℓ1 , . . . , ℓL }

(10)

This offers maximal adaptation capacity but requires careful learning rate control to avoid overfitting and catastrophic forgetting. Augmentation remains disabled, and we rely on layer-wise learning rate decay.

Layer-Wise Learning Rate Decay We apply differentiated learning rates across components to stabilize training across scenarios. Scenario B (3-tier schedule). We use: ηhead > ηadapter > ηenc (11) This accelerates learning for randomly initialized modules while constraining updates to pre-trained layers. Scenario C (depth-based decay). We apply depthdependent decay [Howard and Ruder, 2018]: ηℓj = ηbase · ξ L−j , ξ ∈ (0, 1) (12) Earlier layers receive smaller updates than later layers, reflecting their more general role in encoding ECG morphology. Overall Training Procedure Algorithm 1 summarizes the complete ProFine training loop shared across scenarios, including augmentation only in Scenario A, focal loss optimization, and early stopping based on validation macro-F1. Algorithm 1 CogAdapt: ProFine Training Procedure Require: Pre-trained encoder Eθ , adapter Aϕ , head gψ , target dataset DT Require: Scenario s ∈ {A, B, C}, epochs Es , learning rates {η} 1: Initialize ϕ with Kaiming initialization; initialize ψ randomly 2: Set Θstrain according to chosen scenario (Eq. (8)–(10)) 3: Compute class weights {αc } from training set (Eq. (14)) 4: for epoch = 1 to Es do 5: for batch (X, y) ∼ Dtrain do 6: if s = A then 7: X ← Augment(X) ▷ Augment only in Scenario A 8: end if 9: x12 ← Aϕ (X) ▷ Eq. (3) 10: z ← Pool(Eθ (x12 )) ▷ Eq. (5)–(6) 11: p ← softmax(gψ (z)) ▷ Eq. (7) 12: L ← Lfocal (p, y; α, γ) ▷ Eq. (13) 13: Update Θstrain via AdamW with layer-wise LR decay (Eq. (11)–(12)) 14: end for 15: Evaluate on Dval ; early stop if no improvement in macro-F1 16: end for 17: return Best model checkpoint by validation macro-F1

3.7

Classification Head and Training Objective

Classification Head Architecture The classification head gψ is a lightweight convolutional network operating on pooled encoder features. It applies three sequential 1D convolutional blocks: first projecting R3 → R64 with kernel size 1, followed by BatchNorm1d and ReLU; then R64 → R64 with the same normalization and activation; and finally R64 → R5 with kernel size 1 and bias to produce logits.

Focal Loss with Automatic Rebalancing To handle class imbalance, we minimize a weighted focal loss. Let yk ∈ {0, 1} and pk = [pk,0 , pk,1 ]⊤ denote the true label and predicted probabilities. The per-window loss is: Lk := −αyk (1 − pt )γ log(pt ),

pt = pk,yk

(13)

where γ down-weights easy examples. Automatic class weighting. Weights are computed per fold from training label counts (n0 , n1 ): n0 + n 1 αc = , c ∈ {0, 1} (14) 2nc We monitor validation macro-F1 for early stopping and checkpoint selection to reduce majority-class collapse under imbalance. We note that Scenario A employs focal loss (γ=2) to regularize the randomly initialized adapter and head, whereas Scenarios B and C use standard cross-entropy, as encoder fine-tuning provides sufficient model capacity without additional focal weighting.

4

Experiments

4.1

Experimental Setup

Datasets. We evaluate on two public cognitive load datasets: CLARE (20 subjects performing cognitive tasks) and CLDrive (21 subjects in simulated driving). Both provide 3-lead wearable ECG with self-reported cognitive load labels at 10second intervals (1-9 scale), binarized into low (1-4) versus high (5-9) classes. Preprocessing. Signals are resampled to 500 Hz and bandpass filtered (0.5-40 Hz). We extract 5-second windows with 50% overlap for training and non-overlapping windows for testing. Details are in Appendix C. Evaluation Protocol. We use two cross-validation strategies: (i) 10-fold stratified (K-Fold) and (ii) leave-one-subjectout (LOSO). Early stopping with 10-epoch patience monitors macro-F1 for LOSO and AUROC for K-Fold. Baseline Methods. We compare against ECG-LightCNN and a Transformer encoder trained from scratch, both using identical preprocessing and evaluation protocols. We compare against training-from-scratch baselines using identical 3-lead inputs to isolate transfer learning efficacy. Other baselines either process different inputs, lack published 3-lead protocols, or would necessitate designing custom adapters for each method—conflating adapter architecture effects with pre-training benefits.

4.2

Experiment Analysis

Table 1 reports performance under both K-Fold and LOSO evaluation. Across both datasets and protocols, CogAdapt achieves the strongest results, with Scenario C consistently outperforming training-from-scratch baselines. Under LOSO, which reflects the realistic subject-independent setting, CogAdapt (C) improves macro-F1 from 0.514 to 0.626 on CLARE and from 0.607 to 0.768 on CL-Drive, demonstrating substantially better generalization. Notably, AUROC gains are also pronounced under LOSO (e.g., 0.539 → 0.799 on CLARE), indicating more reliable ranking of cognitiveload states beyond threshold-dependent accuracy.

Under K-Fold validation, which provides a less stringent setting with partial within-subject overlap, CogAdapt (C) still delivers consistent improvements, achieving the best macroF1 on both CLARE (0.785) and CL-Drive (0.860). These results confirm that adapting a clinical-scale ECG foundation encoder is beneficial not only for within-dataset performance, but more importantly for robust cross-subject transfer—the primary bottleneck in wearable cognitive-load modeling.

4.3

Effect of Progressive Fine-Tuning (ProFine Scenarios)

Table 1 also highlights the role of progressive fine-tuning. Scenario A (frozen encoder) provides a strong baseline by leveraging fixed foundation features, but remains limited under LOSO due to insufficient task alignment. Allowing limited encoder adaptation in Scenario B yields consistent gains, particularly in LOSO macro-F1 (e.g., CL-Drive: 0.578 → 0.704), suggesting that unfreezing high-level layers helps resolve representation drift between clinical pre-training and cognitive-load targets. Scenario C achieves the best performance across all settings, indicating that end-to-end adaptation offers the highest capacity to capture subtle autonomic modulations associated with cognitive load, while the progressive design mitigates overfitting and catastrophic forgetting. Beyond aggregate scores, the per-subject analyses (Appendix, Figures 6–9) provide additional evidence that ProFine improves both central tendency and robustness: performance gains are observed for the majority of held-out subjects, and the distribution becomes more favorable compared with from-scratch baselines, supporting more reliable deployment in real-world settings.

4.4

Lead Reconstruction Quality

To evaluate LeadBridge’s effectiveness in transforming 3lead wearable signals into clinically-consistent 12-lead representations, we conducted quantitative reconstruction experiments on the PTB-XL dataset. Figure 4 visualizes the reconstruction quality across all 12 leads, while Table 2 provides detailed metrics for the challenging precordial leads (V2– V6). Perfect reconstruction of standard leads. As shown in Figure 4, LeadBridge achieves near-perfect reconstruction for the first seven leads (I, II, III, aVR, aVL, aVF, and V1), with the predicted waveforms (red dashed) exhibiting almost complete overlap with ground-truth signals (blue solid). This demonstrates that LeadBridge successfully learns the deterministic relationships governing limb and augmented limb leads, while maintaining anatomical consistency. Superior performance on precordial leads. Table 2 compares LeadBridge against two established baselines: the Dower inverse transform [Dower et al., 1980] and linear regression [Obianom et al., 2025]. LeadBridge achieves the lowest RMSE for four out of five precordial leads (V2, V3, V4, V5), with particularly notable improvements on V2 (175.61 vs. 195.08) and V3 (189.77 vs. 194.52). More importantly, LeadBridge demonstrates substantially higher correlation coefficients across all leads except V6, with V2

CLARE Method

K-Fold Acc

F1

ECG-LightCNN [Bhatti et al., 2024] .774 .726 Transformer [Bhatti et al., 2024] .705 .643 CogAdapt (Scenario A) CogAdapt (Scenario B) CogAdapt (Scenario C)

.742 .717 .762 .736 .813 .785

CL-D RIVE LOSO

AUC Acc

F1

K-Fold AUC Acc

F1

LOSO AUC Acc

F1

AUC

.832 .718

.686 .514 .623 .496

.539 .506

.813 .808 .688 .680

.887 .731

.700 .607 .611 .541

.707 .604

.843 .857 .898

.681 .527 .691 .557 .736 .626

.598 .706 .799

.823 .821 .833 .831 .862 .860

.906 .919 .945

.665 .578 .770 .704 .831 .768

.705 .811 .889

Table 1: Performance on CLARE and CL-Drive under K-Fold and LOSO validation.

Lead Dower Linear Regression RMSE ↓ V2 205.69 195.08 290.39 194.52 V3 V4 270.63 164.72 V5 215.06 143.45 V6 161.40 129.66 C ORRELATION C OEFFICIENT (CC) ↑ V2 0.546 0.546 V3 0.569 0.570 V4 0.662 0.698 0.679 0.694 V5 V6 0.544 0.549

LeadBridge 175.61 189.77 164.02 140.76 135.22 0.669 0.613 0.718 0.717 0.547

Table 2: Performance comparison of ECG lead reconstruction methods across precordial leads V2-V6. Green highlighting indicates best performance for each lead.

showing the largest gain (CC=0.669 vs. 0.546)—a 22.5% relative improvement over both baselines. Challenging nature of precordial reconstruction. While limb leads can be derived through deterministic linear combinations, precordial leads require learning complex spatialtemporal mappings from limited input channels. Figure 4 reveals larger reconstruction errors for V2–V6 (visible gaps between blue and red traces), particularly in peak amplitudes and subtle morphological details. This inherent difficulty explains why even our learned approach achieves moderate correlation coefficients (0.61–0.72 range). Importantly, these reconstruction metrics reflect PTB-XL pre-training alone. During downstream training, LeadBridge is further optimized end-to-end for cognitive load classification to learn to emphasize task-relevant signal components rather than reconstruction of signals only. The classification gains in Table 1 confirm that this task-adapted transformation effectively bridges wearable inputs to foundation model representations. Nevertheless, LeadBridge’s consistent superiority over fixed transforms validates the benefit of data-driven adaptation for handling non-standard electrode placements typical in wearable devices. The reconstruction quality directly impacts downstream cognitive load classification: by providing anatomicallyconsistent 12-lead inputs to ECG-FM, LeadBridge enables effective transfer of clinical pre-training to wearable sensing scenarios despite the fundamental sensor configuration mis-

Figure 4: ECG lead reconstruction on PTB-XL. Reconstructions of all 12 leads from a 3-lead wearable input. The blue solid waveform denotes the ground-truth signal, while the red dashed waveform indicates the predicted (reconstructed) lead. The first seven leads (I, II, III, aVR, aVL, aVF, and V1) are recovered almost perfectly with near-complete overlap, whereas larger deviations are observed in the more challenging precordial leads (V2–V6).

match.

5

Conclusion

We propose CogAdapt, a framework that adapts clinical ECG foundation models to wearable cognitive load assessment. CogAdapt addresses two key challenges through LeadBridge, a learnable adapter that transforms 3-lead wearable signals into anatomically-consistent 12-lead representations, and ProFine, a progressive fine-tuning strategy that prevents catastrophic forgetting while enabling task adaptation. Experiments on CLARE and CL-Drive datasets under leaveone-subject-out evaluation show that CogAdapt substantially outperforms training-from-scratch baselines, achieving

macro-F1 scores of 0.626 and 0.768 respectively. Our results demonstrate that transfer learning from large-scale clinical ECG data can effectively address the limited labeled data and poor cross-subject generalization challenges in wearable cognitive load assessment. Future work will explore extending this approach to other physiological signals and investigating personalized adaptation with minimal calibration data.

References [Abbaspourazad et al., 2024] Salar Abbaspourazad, Oussama Elachqar, Andrew C. Miller, Saba Emrani, Udhyakumar Nallasamy, and Ian Shapiro. Large-scale training of foundation models for wearable biosignals. In ICLR, 2024. [Angkan et al., 2024] Prithila Angkan, Behnam Behinaein, Zunayed Mahmud, Anubhav Bhatti, Dirk Rodenburg, Paul Hungler, and Ali Etemad. Multimodal brain–computer interface for in-vehicle driver cognitive load measurement: Dataset and baselines. IEEE Transactions on Intelligent Transportation Systems, 25(6):5949–5964, 2024. [Ayres et al., 2021] Paul Ayres, Joy Yeonjoo Lee, Fred Paas, and Jeroen J. G. Van Merrienboer. The validity of physiological measures to identify differences in intrinsic cognitive load. Frontiers in Psychology, 12:702538, 2021. [Bhatti et al., 2024] Anubhav Bhatti, Prithila Angkan, Behnam Behinaein, Zunayed Mahmud, D. Rodenburg, Heather Braund, P. Mclellan, A. Ruberto, Geoffery Harrison, Daryl Wilson, A. Szulewski, Dan Howes, A. Etemad, and Paul Hungler. Clare: Cognitive load assessment in realtime with multimodal data. IEEE Transactions on Cognitive and Developmental Systems, 2024. [Chen et al., 2020] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 1597–1607. PMLR, 13–18 Jul 2020. [Dower et al., 1980] Gordon E Dower, H Bastos Machado, and John A Osborne. On deriving the electrocardiogram from vectorcardiographic leads. Clinical cardiology, 3(2):87–95, 1980. [Feild et al., 2008] Dirk Q. Feild, Sophia H. Zhou, Eric D. Helfenbein, Richard E. Gregg, and James M. Lindauer. Technical challenges and future directions in lead reconstruction for reduced-lead systems. Journal of Electrocardiology, 41(6):466–473, 2008. [Gjoreski et al., 2020] Martin Gjoreski, Tine Kolenik, Timotej Knez, Mitja Luštrek, Matjaž Gams, Hristijan Gjoreski, and Veljko Pejović. Datasets for cognitive load inference using wearable sensors and psychological traits. Applied Sciences, 10(11):3843, 2020. [Gkintoni et al., 2025] Evgenia Gkintoni, Hera Antonopoulou, Andrew Sortwell, and Constantinos Halkiopoulos. Challenging cognitive load theory: The role of educational neuroscience and artificial intelligence in redefining learning efficacy. Brain Sciences, 15:203, 02 2025. [Gu et al., 2025] Xiao Gu, Yuxuan Shu, Jinpei Han, Yuxuan Liu, Zhangdaihong Liu, James Anibal, Veer Sangha, Edward Phillips, Bradley Segal, Yuxuan Liu, Hang Yuan, Fenglin Liu, Kim Branson, Patrick Schwab, Danielle Belgrave, Lei Clifton, Dimitris Spathis, Vasileios Lampos, A Aldo Faisal, and David A Clifton. Foundation models for biosignals: A survey, August 2025. TechRxiv Preprint.

[Han et al., 2024] Yu Han, Xiaofeng Liu, Xiang Zhang, and Cheng Ding. Foundation models in electrocardiogram: A review. arXiv preprint arXiv:2410.19877, 2024. [Han et al., 2025] Yu Han, Vittorio Murino, Xiaofeng Liu, Xiang Zhang, and Cheng Ding. A systematic review on foundation models for electrocardiogram analysis: Initial strides and expansive horizons, 2025. [Hassannia and Sameni, 2025] Mohammadsina Hassannia and Reza Sameni. A neural network architecture for ecg lead reconstruction: Separating shared and lead-specific ecg characteristics. Computing in Cardiology Conference (CinC), 2025. [He et al., 2015] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026–1034, 2015. [Howard and Ruder, 2018] Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 328– 339, 2018. [Hughes et al., 2019] Ashley M. Hughes, Gabriella M. Hancock, Shannon L. Marlow, Kimberly Stowers, and Eduardo Salas. Cardiac measures of cognitive workload: a meta-analysis. Human factors, 61(3):393–414, 2019. [Ioffe, 2015] Sergey Ioffe. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. [Kalatzis and others, 2021] Ilias Kalatzis et al. A database for cognitive workload classification using electrocardiogram and respiration signal, 2021. Dataset description; please fill in the correct journal/venue, volume, pages, and DOI from the official publication. [Kim et al., 2022] S. Kim, S. Chon, J. K. Kim, J. Kim, Y. Gil, and S. Jung. Lightweight convolutional neural network for real-time arrhythmia classification on low-power wearable electrocardiograph. In Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBS), pages 1915–1918, July 2022. [Kiyasseh et al., 2021] Dani Kiyasseh, Tingting Zhu, and David A Clifton. Clocs: Contrastive learning of cardiac signals across space, time, and patients. In International Conference on Machine Learning, pages 5606– 5615. PMLR, 2021. [Kligfield et al., 2007] Paul Kligfield, Leonard S Gettes, James J Bailey, Rory Childers, Barbara J Deal, E William Hancock, Gerard Van Herpen, Jan A Kors, Peter Macfarlane, David M Mirvis, et al. Recommendations for the standardization and interpretation of the electrocardiogram: part i: the electrocardiogram and its technology a scientific statement from the american heart association electrocardiography and arrhythmias committee, council on clinical cardiology; the american college of cardiology foundation; and the heart rhythm society endorsed

by the international society for computerized electrocardiology. Journal of the American College of Cardiology, 49(10):1109–1127, 2007. [Kosch et al., 2023] Thomas Kosch, Jakob Karolus, Johannes Zagermann, Harald Reiterer, Albrecht Schmidt, and Paweł W. Woźniak. A survey on measuring cognitive workload in human-computer interaction. ACM Comput. Surv., 55(13s), July 2023. [Kundu, 2022] Subhasis Kundu. Hci-driven emotionadaptive uis for cognitive efficiency: Real-time adaptation to fatigue, focus, and emotional expression. International Scientific Journal of Engineering and Management, 01:1– 7, 09 2022. [Lawhern et al., 2013] Vernon Lawhern, W. David Hairston, and Kay Robbins. Detect: A matlab toolbox for event detection and identification in time series, with applications to artifact detection in eeg signals. PLOS ONE, 8(4):1–13, 04 2013. [Markova and others, 2019] Veselina Markova et al. CLAS: A database for cognitive load, affect, and stress, 2019. Dataset description; please insert the correct venue/journal information from the official publication. [McKeen et al., 2025] Kaden McKeen, Sameer Masood, Augustin Toma, Barry Rubin, and Bo Wang. Ecg-fm: An open electrocardiogram foundation model. JAMIA open, 8(5):ooaf122, 2025. [Moeinzadeh et al., 2018] Hossein Moeinzadeh, Gaetano Gargiulo, Paolo Bifulco, Mario Cesarelli, A O’Loughlin, MI Shugman, and Aravinda Thiagalingam. A modern wilson’s central terminal electrocardiography database. Heart, Lung and Circulation, 27:S293–S294, 2018. [Mvuh et al., 2024] Franck Lino Mvuh, Claude Odile Vanessa Ebode Ko’a, and Bertrand Bodo. Multichannel high noise level ecg denoising based on adversarial deep learning. Scientific Reports, 14(1):801, 2024. [Naismith and Cavalcanti, 2015] Laura M. Naismith and Rodrigo B. Cavalcanti. Validity of cognitive load measures in simulation-based training: a systematic review. Academic Medicine, 90(11):S24–S35, Nov 2015. [Obianom et al., 2025] Ekenedirichukwu N Obianom, G André Ng, and Xin Li. Reconstruction of 12-lead ecg: a review of algorithms. Frontiers in Physiology, 16:1532284, 2025. [Potharaju et al., 2025] Saiprasad Potharaju, K. Veera Raghavulu, N. S. Pradeep Kumar, Suresh Salendra, Swapnali N. Tambe, and M. V. V. Prasad Kantipudi. Multi-resolution hybrid sliding window approach for ecg arrhythmia detection using deep learning models. Discover Applied Sciences, 7(12):1412, 2025. [Rahman et al., 2022] Saifur Rahman, Chandan Karmakar, Iynkaran Natgunanathan, John Yearwood, and Marimuthu Palaniswami. Robustness of electrocardiogram signal quality indices. Journal of The Royal Society Interface, 19(189):20220012, 04 2022.

[Raza et al., 2024] Muhammad Shahzad Raza, Muhammad Murtaza, C. T. Cheng, and Muhammad Asif Khan. Systematic review of cognitive impairment in drivers through mental workload using physiological measures of heart rate variability. Frontiers in Computational Neuroscience, 18:1475530, 2024. [Satija et al., 2018] U. Satija, B. Ramkumar, and M. S. Manikandan. A review of signal processing techniques for electrocardiogram signal quality assessment. IEEE Reviews in Biomedical Engineering, 11:176–191, 2018. [Shang et al., 2021] Zuogang Shang, Zhibin Zhao, Hui Fang, Samuel Relton, Darcy Murphy, Zoe Hancox, Ruqiang Yan, and David Wong. Deep discriminative domain generalization with adversarial feature learning for classifying ecg signals. In 2021 Computing in Cardiology (CinC), volume 48, pages 1–4, Sep. 2021. [Solhjoo et al., 2019] S. Solhjoo, M. C. Haigney, E. McBee, and S. J. Durning. Heart rate and heart rate variability correlate with clinical reasoning performance and selfreported measures of cognitive load. Scientific Reports, 9(1):1–9, 2019. [Suryani et al., 2024] Mira Suryani, Harry Budi Santoso, Martin Schrepp, Rizal Fathoni Aji, Setiawan Hadi, Dana Indra Sensuse, Ryan Randy Suryono, et al. Role, methodology, and measurement of cognitive load in computer science and information systems research. IEEE Access, 2024. [Vozda and Cerny, 2015] M. Vozda and M. Cerny. Methods for derivation of orthogonal leads from 12-lead electrocardiogram: A review. Biomedical Signal Processing and Control, 19:23–34, 2015. [Wagner et al., 2020] Patrick Wagner, Nils Strodthoff, RalfDieter Bousseljot, Dieter Kreiseler, Fatima I Lunze, Wojciech Samek, and Tobias Schaeffter. Ptb-xl, a large publicly available electrocardiography dataset. Scientific Data, 7(1):1–15, 2020. [Yang et al., 2022] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. Simper: Simple self-supervised learning of periodic targets. arXiv preprint arXiv:2210.03115, 2022. [Yang et al., 2023a] Chaoqi Yang, M Westover, and Jimeng Sun. Biot: Biosignal transformer for cross-data learning in the wild. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 78240– 78260. Curran Associates, Inc., 2023. [Yang et al., 2023b] Chia-Yen Yang, Pin-Chen Chen, and Wen-Chen Huang. Cross-domain transfer of eeg to eeg or ecg learning for cnn classification models. Sensors, 23(5), 2023.

Appendix A

Baselines

We compare our approach against two baseline methods for cognitive load classification from ECG signals: ECG-LightCNN represents the family of lightweight convolutional architectures for end-to-end ECG analysis [Bhatti et al., 2024]. This baseline uses a shallow CNN inspired by VGG networks with multiple convolutional and max-pooling layers, followed by fully connected layers for binary classification. The model learns features directly from raw ECG waveforms without any pre-training, making it representative of task-specific supervised learning approaches. We follow the architecture specification from the CLARE dataset paper, which uses 3-4 convolutional blocks with ReLU activations and dropout for regularization. Transformer follows the architecture proposed in the CLARE dataset paper [Bhatti et al., 2024]. This baseline uses multi-head self-attention mechanisms to model temporal dependencies in ECG sequences. The model applies positional encoding to raw ECG windows and processes them through stacked transformer encoder blocks. Like the CNN baseline, this transformer is trained from scratch on the cognitive load task without leveraging large-scale pre-training. The architecture uses temporal mean pooling over the sequence dimension before passing features to a classification head for binary cognitive load prediction. All baselines are evaluated using the same LOSO crossvalidation protocol as our main approach to ensure fair comparison of subject-independent generalization performance. This appendix provides comprehensive technical details supporting the methodology and experiments presented in the main paper.

B

Dataset Characteristics and Limitations

This section provides comprehensive analysis of publicly available datasets for cognitive load assessment, examining their suitability for real-time, subject-independent modeling.

B.1

modeling, though inter-subject variability remains substantial. The MATB-II (Multi-Attribute Task Battery II) protocol requires participants to simultaneously manage four subtasks: system monitoring (detecting indicator changes), tracking (controlling a cursor), communications (responding to radio calls), and resource management (maintaining fuel levels). Task difficulty is manipulated by varying subtask demands across blocks, inducing different cognitive load levels. The dataset’s strength lies in its ecological validity—MATB-II simulates realistic multitasking scenarios—and temporal label density. However, the limited number of subjects (20 in public release) and strong individual differences in physiological responses create challenges for subject-independent generalization. CL-Drive (Cognitive Load in Driving). CL-Drive captures multimodal signals (EEG, ECG, EDA, and eye tracking) from 21 participants in a driving simulator under scenarios of varying complexity [Angkan et al., 2024]. Drivers verbalize their perceived workload at 10-second intervals using a 9-point PAAS scale, providing temporal alignment similar to CLARE. Each participant performed driving tasks across 9 complexity levels, with each level lasting 3 minutes. The dataset includes diverse driving conditions (e.g., highway, urban, adverse weather) designed to elicit different levels of cognitive demand. ECG is recorded at 256 Hz using a chestworn sensor. The dataset enables evaluation of cognitive load models in safety-critical contexts where real-time assessment could support adaptive driver assistance systems. Driving scenarios range from simple highway cruising (low load) to complex urban navigation with pedestrians, traffic lights, and lane changes (high load). Weather conditions (rain, fog) and traffic density are systematically varied to manipulate cognitive demand. The simulator-to-real domain gap remains a limitation—drivers’ physiological responses in simulated environments may differ from real driving—but the dataset provides valuable benchmarks for comparing cognitive load inference methods across different task domains.

Datasets Used in This Work

We evaluate our approach on two publicly available cognitive load datasets that provide fine-grained, time-aligned labels suitable for real-time assessment. Both datasets include ECG among their multimodal physiological recordings and report subject-independent evaluation metrics. CLARE (Cognitive Load Assessment in REaltime). CLARE contains multimodal physiological signals (ECG, EDA, EEG, and eye gaze) from 24 participants performing MATB-II multitasking protocols with varying cognitive load across blocks [Bhatti et al., 2024]. Participants provide insitu verbal ratings of their perceived cognitive load every 10 seconds on a 9-point Likert scale (1-9), enabling dense time-aligned supervision. Each participant completed four 9minute sessions, with each session comprising nine randomly selected 1-minute complexity levels. The public release contains 20 subjects with 3-lead wearable ECG recorded at 256 Hz. This dataset addresses key limitations of prior work by providing fine-grained labels that support real-time predictive

B.2

Taxonomy of Cognitive Load Datasets

Table 3 presents a systematic comparison of cognitive load datasets across five dimensions: physiological modalities, labeling granularity, participant scale, task paradigms, and key limitations for real-time inference.

B.3

Labeling Paradigms and Their Impact on Modeling

Retrospective vs. In-Situ Labeling. Early datasets (CogLoad, Snake) rely on retrospective labels collected after task completion, introducing memory biases and temporal misalignment. Participants must recall mental effort experienced minutes earlier, producing labels that reflect aggregate impressions rather than momentary states. This limits model training to coarse block-level predictions unsuitable for realtime monitoring. In contrast, CLARE and CL-Drive employ in-situ labeling where participants report load every 10 seconds during task

Dataset

Modalities

Label scale

CogLoad [Gjoreski et al., 2020] Snake [Gjoreski et al., 2020]

EDA, ACC, ST, RR (IBI)

Post-task NASATLX (per task trial)

23

EDA, ACC, ST, RR (IBI)

Post-task self-rating (per game level)

23

CLAS [Markova et al., 2019]

ECG, PPG, EDA, ACC

Task-block labels + post-task ratings (often categorical high/low)

62

EEG, ECG, EDA, eye tracking

In-situ ratings every 10 s / 9-point PAAS

21

Driving simulator (9 complexity levels, 3 min each)

ECG, EDA, EEG, eye gaze

In-situ verbal ratings every 10 s / Likert 1– 9

24†

MATB-II multitask (four 9 min sessions; nine 1 min complexity levels)

CL-Drive [Angkan al., 2024]

et

CLARE [Bhatti et al., 2024]

cadence

/

#Subj.

Stimuli / tasks

Key limitations

6 PC-based cognitive tasks (e.g., nback) Smartphone Snake game (easy, medium, hard; ≥2 min each) Math/Logic/Stroop (3 interactive); Image/Video (2 perceptive)

Retrospective, coarse (one label per multi-minute trial); no raw ECG (only RR/IBI); limited within-trial dynamics. Retrospective labels; level-based supervision; no raw ECG.

Labels are block-level (not dense time series); mixed conditions (load + affect/stress) can confound ”pure” load; not designed for fine-grained sequence learning. Dense, time-aligned self-report; simulator-to-real domain gap; subjectindependent evaluation (e.g., LOSO) reported in benchmarks. Dense labels enable real-time learning; but limited hours/participants; strong inter-subject variability and lower LOSO performance vs random splits.

Table 3: Systematic comparison of public cognitive load datasets. † The CLARE paper reports 24 participants, while the public release contains 20 subjects.

performance. This temporal alignment enables window-wise supervision: each 5-10 second ECG segment can be paired with a contemporaneous load label. However, self-reporting introduces task interruption and potential reactivity effects— the act of reporting may itself alter cognitive state. Label Granularity and Frequency. Label frequency directly impacts model temporal resolution. Block-level labels (CLAS: one label per task block) support only coarse-grained classification, missing within-trial dynamics critical for adaptive systems. The 10-second cadence in CLARE/CL-Drive enables real-time inference but still averages over brief cognitive fluctuations. Finer granularity remains an open challenge given the intrusiveness of frequent self-reports.

B.4

Scale and Generalization Challenges

Participant Counts and Data Volume. Even ”denselabel” datasets provide limited training data: CLARE’s 20 subjects × 36 minutes/subject ≈ 12 hours total; CL-Drive’s 21 subjects × 27 minutes/subject ≈ 9.5 hours total. Compare this to ECG-FM’s pre-training corpus: 1.5 million ECGs from large-scale public datasets (MIMIC-IV-ECG, PhysioNet 2021). This orders of magnitude gap motivates transfer learning—supervised training from scratch cannot match foundation models’ generalization capacity built on extensive pre-training. Inter-Subject Variability. Subject-independent evaluation reveals substantial performance drops. Bhatti et al. report CLARE CNN accuracy: 78.4% (k-fold) → 63.3% (LOSO), a 15.1pp degradation [Bhatti et al., 2024]. This gap quantifies individual differences in baseline physiology, sensor coupling, and cognitive strategy. Foundation models pre-trained

on diverse populations may learn more invariant features than task-specific models trained on less than 25 subjects.

B.5

Dataset Access and Availability

Both datasets are publicly available for research purposes: • CLARE: https://borealisdata.ca/dataset.xhtml? persistentId=doi:10.5683/SP3/H0AELT • CL-Drive: https://borealisdata.ca/dataset.xhtml? persistentId=doi:10.5683/SP3/JJ2YZZ These datasets have become standard benchmarks for evaluating cognitive load inference methods, enabling direct comparison with prior work while promoting reproducibility in the field.

C

Signal Processing and Preprocessing Strategies

C.1

Lead Reconstruction: Theoretical Foundation and Implementation Details

This part provides theoretical background and implementation details for the LeadBridge adapter and Wilson Central Terminal (WCT) re-referencing described in the main paper. The standard clinical 12-lead ECG comprises three limb leads (I, II, III), three augmented limb leads (aVR, aVL, aVF) derived via Goldberger’s equations, and six precordial leads (V1 –V6 ). This requires 10 electrodes for comprehensive cardiac activity assessment.

The wearable device provides three leads: Lead I = VLA − VRA , Lead II = VLL − VRA ,

(15) (16)

Lead Vx = Vxchest − VRL ,

(17)

where LA, RA, LL, and RL denote left arm, right arm, left leg, and right leg electrodes, respectively. This configuration differs from clinical standards in chest lead positioning and referencing (RL instead of WCT). The WCT is defined as: VRA + VLA + VLL . (18) WCT = 3 A clinical chest lead is: VxWCT = Vx − WCT.

(19)

Given the wearable’s VxRL = Vx − VRL and assuming RL ≈ RA, the conversion is: VxWCT = VxRL +

I + II . 3

(20)

This yields the re-referenced input xref = [I, II, VxWCT ]⊤ . Re-referencing is applied before window-level normalization to preserve voltage relationships. PTB-XL provides 12-lead ECGs referenced to WCT. To simulate 3-lead inputs: - Limb leads: Lead I and II are directly available. - Chest lead: Select V2 as x; simulate V2RL via: I + II . (21) V2RL = V2WCT − 3 RL This yields [I, II, V2 ] with ground-truth 12-lead targets. Filter for normal sinus rhythm (SCP code NORM ≥ 80% confidence), yielding ∼9,500 recordings. Use 10second windows with 50% overlap; split 80/20 for training/validation. LeadBridge Aϕ : R3×T → R12×T uses a linear projection: Aϕ (x) := σ(BatchNorm(W x)),

W ∈ R12×3 ,

(22)

• Subject-wise: Statistics computed per subject. Creates leakage in LOSO evaluation, as test-time normalization requires statistics from the held-out subject. • Window-wise: Statistics computed independently per window. Removes baseline and amplitude variations within each segment, preventing subject-specific information leakage.

C.3

Bandpass Filtering and Artifact Handling

Filter Design. We apply a zero-phase 4th-order Butterworth bandpass filter (0.5–40 Hz) using forward-backward filtering. This removes: (i) baseline wander (<0.5 Hz) from respiration and motion artifacts, and (ii) high-frequency noise (>40 Hz) from muscle activity, powerline interference, and instrumentation. Zero-phase filtering ensures no temporal distortion of cardiac waveforms (P-wave, QRS complex, Twave morphology preserved). Handling Missing and Degenerate Data. servative data cleaning:

1. Missing samples within a lead: Zero-fill individual NaN values (rare, <0.1% of samples). 2. Constant/flat lead: If a lead exhibits zero variance (e.g., electrode disconnection), set that lead to zero to avoid degenerate normalization. 3. Entirely missing rows: Discard data points where all ECG leads are simultaneously missing. This conservative approach prioritizes signal quality over data quantity.

C.4

Data Augmentation

During scenario A training (adapter and classification head only), we apply stochastic augmentations to improve generalization. Each augmentation is applied independently with probability p = 0.5:

where σ is ReLU. Implemented as Conv1D (kernel size 1, stride 1), BatchNorm1D, and ReLU. Batch normalization aligns distributions with the foundation model’s inputs. Initialize W with Kaiming initialization; BatchNorm affine parameters to identity. Pre-train on PTB-XL using MSE loss for 3→12 lead reconstruction. Freeze weights during downstream cognitive load training. Limb leads III, aVR, aVL, aVF use deterministic relationships (Einthoven’s Law, Goldberger’s equations). Precordial leads V1 , V3 –V6 are learned via the projection, accommodating non-standard placements without hard-coded constraints.

Amplitude Jitter.

C.2

Temporal Shift.

Normalization Strategies for Subject-Independent Modeling

Normalization Schemes. Three normalization strategies exist for ECG signals: • Global: Mean µ and standard deviation σ computed over the entire training set. Fails to account for intersubject amplitude differences and requires storing population statistics.

We employ con-

Per-lead multiplicative scaling:

x′c = xc · (1 + α · zc ),

zc ∼ N (0, 1)

(23)

where α = 0.1 and c indexes the lead. This simulates electrode impedance variations. Additive Gaussian Noise. x′ = x + η,

η ∼ N (0, σ 2 I)

(24)

with σ = 0.02 (applied after z-score normalization). This improves robustness to measurement noise. Circular shift by δ samples:

x′ [t] = x[(t + δ)

mod T ],

δ ∼ U{−50, 50}

(25)

At 500 Hz, this corresponds to ±100 ms temporal jitter, simulating window alignment variations. Augmentations are disabled during scenarios B and C (encoder fine-tuning) to avoid disrupting pre-trained feature representations.

C.5

Windowing and Resampling

Window Extraction. We extract 5-second windows using timestamp-based segmentation to respect data gaps (e.g., from sensor disconnections). For training data, we use 50% overlap (2.5-second stride) to increase sample count; for validation/test, we use 0% overlap to prevent temporal leakage between evaluation windows. Resampling. All windows are resampled to 500 Hz using polyphase FIR resampling, yielding fixed-length outputs of T = 2500 samples per window. Resampling is performed before filtering to ensure consistent filter behavior across varying source sampling rates. Label Alignment. Self-reported cognitive load labels are collected at 10-second intervals. We assign each 5-second window the label corresponding to the 10-second epoch containing the window’s temporal center:   tcenter (w) − t0 label(w) = labels (26) ∆tlabel

Scenario A: Feature Evaluation. Convergence is rapid (∼15 epochs) due to the small trainable parameter count (≈199K). The frozen encoder provides stable features, allowing the adapter and head to converge quickly. Scenario C: Full Fine-Tuning. Early stopping with patience=10 is critical; typical best epochs occur between 3–8. The large parameter space and small dataset make the model prone to overfitting beyond this range.

D.4

Plasticity–Stability Trade-off

The progression A → B → C reveals a plasticity–stability trade-off with distinct patterns across evaluation settings, as illustrated in Figure 5 and Figure 6.

where t0 is the experiment start time and ∆tlabel = 10 s.

D

Progressive Fine-Tuning: Theoretical Motivation

This section provides the theoretical foundation and mathematical details for the progressive unfreezing strategy introduced in Section 3.5 of the main paper.

D.1

Catastrophic Forgetting in Transfer Learning

When fine-tuning all parameters from a randomly initialized task head: θt+1 = θt − η∇θ Ltask (θt ) (27) large gradients from the randomly initialized head propagate through the encoder, causing drastic weight changes in early layers. Pre-trained features encoding general cardiac patterns are overwritten to fit the limited downstream data, leading to overfitting and degraded generalization.

D.2

Tiered Learning Rate Decay

To mitigate forgetting, we apply tier-wise learning rate decay, assigning progressively smaller learning rates to deeper (earlier) layers. For ECG-FM’s L = 12 Transformer layers, we partition the encoder into three groups: Gtop = {ℓ9 , . . . , ℓ12 },

(28)

Gmid = {ℓ5 , . . . , ℓ8 }, (29) Gbot = {ℓ1 , . . . , ℓ4 } (30) where ηtop > ηmid > ηbot . Lower layers capture low-level signal structure (QRS morphology, P-wave shapes) that generalizes across tasks and patient populations. Upper layers encode progressively more task-specific features. Smaller learning rates for lower layers preserve general representations while allowing upper layers to adapt to cognitive-load-specific patterns.

D.3

Convergence Characteristics Across Scenarios

The progressive unfreezing strategy described in Section 3.5 exhibits distinct convergence behaviors:

Figure 5: Plasticity–stability trade-off for CLARE dataset across fine-tuning scenarios (A: frozen encoder, B: top-layer adaptation, C: full fine-tuning). Error bars represent standard deviation across folds (K-Fold) or subjects (LOSO). While K-Fold shows controlled variance growth, LOSO reveals substantially increasing variability with greater model plasticity.

K-Fold Cross-Validation. Under K-fold evaluation (within-subject data splitting), we observe a consistent trade-off across both datasets: • Scenario A: Highest stability with lowest standard deviations (CLARE: ±0.010; CL-DRIVE: ±0.018), but limited adaptation yields baseline performance. • Scenario B: Modest performance gains with marginally increased variance (CLARE: ±0.012; CL-DRIVE: ±0.019). Top-layer adaptation captures task-specific patterns while preserving low-level ECG representations. • Scenario C: Maximum plasticity achieves highest performance at the cost of elevated variance (CLARE: ±0.034; CL-DRIVE: ±0.025), indicating greater sensitivity to fold composition. Leave-One-Subject-Out (LOSO). Under the more challenging LOSO regime, the trade-off manifests differently between the two datasets (Fig. 5 vs. Fig. 6): • CLARE: Progressive increase in both performance and variance (±0.052 → ±0.085 → ±0.119). The widening error bars across scenarios A, B, and C indicate that greater model flexibility improves cognitive load detection but reduces cross-subject generalization stability.

TP c TP c + FN c The class-wise F1 score is then computed as: Rc =

F 1c =

2Pc Rc . Pc + R c

(32)

(33)

Macro F1 averages uniformly across both classes, ensuring equal weight regardless of class imbalance: 1

F 1macro =

Figure 6: Plasticity–stability trade-off for CL-DRIVE dataset across fine-tuning scenarios. Note the non-monotonic variance pattern in LOSO: Scenario A exhibits high instability (±0.149), which improves in Scenario B (±0.099), suggesting that frozen ECG representations are poorly aligned with driving-specific cognitive load patterns.

• CL-DRIVE: Scenario A exhibits surprisingly high variance (±0.149), which decreases substantially in Scenario B (±0.099) before moderately increasing in Scenario C (±0.117). This non-monotonic pattern, clearly visible in Fig. 6, suggests that frozen representations may be misaligned with driving-specific cognitive patterns, and that controlled adaptation (Scenario B) stabilizes predictions while substantially improving performance. Practical Implications. As shown in both figures, Scenario C consistently achieves the highest performance when within-subject data is available (K-Fold), making it suitable for personalized applications. For cross-subject generalization (LOSO), the optimal choice is task-dependent: CLDRIVE benefits most from Scenario B, which offers a favorable stability-performance balance (±0.099), while CLARE shows continued performance gains through Scenario C despite increased variability (±0.119). The contrasting variance patterns visible in Figure 5 and Figure 6 highlight the importance of task-specific fine-tuning strategies when adapting foundation models to cognitive state estimation.

E

Evaluation Metrics

Let D = {(xi , yi )}N i=1 denote a test set of N labeled ECG windows, where yi ∈ {0, 1} represents binary cognitive load labels (0 = low, 1 = high) and ŷi is the predicted class. Let p̂i ∈ [0, 1] denote the predicted probability for the positive class. For binary classification, let TP , FP , TN , and FN denote the counts of true positives, false positives, true negatives, and false negatives, respectively.

E.1

Macro F1 Score

TP c TP c + FP c

(31)

(34)

Accuracy

Accuracy measures the proportion of correctly classified windows: N TP + TN 1 X = ⊮[ŷi = yi ] TP + TN + FP + FN N i=1 (35) where 1[·] denotes the indicator function.

Accuracy =

E.3

Area Under the ROC Curve (AUROC)

The area under the receiver operating characteristic curve (AUROC) evaluates discrimination ability across all classification thresholds. For each threshold τ ∈ [0, 1], we define the true positive rate and false positive rate: TPR(τ ) =

|{i : p̂i ≥ τ, yi = 1}| |{i : yi = 1}|

(36)

FPR(τ ) =

|{i : p̂i ≥ τ, yi = 0}| |{i : yi = 0}|

(37)

The AUROC is computed as: Z 1 AUROC = TPR d(FPR)

(38)

0

which is equivalent to the probability that a randomly chosen positive example receives a higher predicted probability than a randomly chosen negative example. In practice, we compute AUROC using the trapezoidal rule over the empirical ROC curve.

F

Hyperparameters

For reproducibility, we provide the complete hyperparameter configurations for our proposed approach. These settings were carefully tuned to ensure robust performance across both CLARE and CL-Drive datasets under K-fold and LOSO evaluation protocols. All experiments were conducted on a single NVIDIA A100 GPU.

F.1

For each class c ∈ {0, 1}, we define precision and recall under the one-vs-rest convention: Pc =

E.2

F 10 + F 11 1X F 1c = 2 c=0 2

Progressive Fine-Tuning Configurations

We implement our approach using PyTorch with automatic mixed precision (AMP) enabled for computational efficiency. The complete hyperparameter settings for each scenario are summarized in Tables 4, 5, 6, and 7.

Hyperparameter K-Fold LOSO Scenario A (Frozen Encoder) Training Epochs 30 30 Batch Size 64 64 LR Head 1e−3 1e−3 LR Adapter 1e−4 1e−4 Loss Function Focal (γ=2) Focal (γ=2) Class Weighting Auto-α Auto-α Data Augmentation Enabled Enabled Scenario B (Partial Unfreeze) Training Epochs 20 20 Batch Size 64 64 LR Head 5e−4 5e−4 LR Adapter 1e−4 1e−4 LR Encoder (Top 2) 1e−5 1e−5 Loss Function Cross-Entropy Cross-Entropy Data Augmentation Disabled Disabled Scenario C (Full Fine-Tuning) Training Epochs 10 10 Batch Size 32 32 LR Head 1e−4 5e−5 LR Adapter 3e−5 1e−5 LR Encoder (Top) 1e−5 5e−6 LR Encoder (Mid) 3e−6 1e−6 LR Encoder (Bottom) 1e−6 5e−7 Loss Function Cross-Entropy Cross-Entropy Data Augmentation Disabled Disabled

Configuration Value Data Augmentation (Scenario A) Amplitude Jitter ±10% (p=0.5) Gaussian Noise σ 0.02 (p=0.5) Time Shift ±50 samp (p=0.5) Data Processing Sampling Rate 500 Hz Resampling Method Polyphase FIR Bandpass Filter 0.5–40 Hz Filter Order 4th Butterworth Filter Type Zero-phase Window Size 5s (2500 samp) Train Stride 2.5s (50% overlap) Test Stride 5s (no overlap) Normalization Per-window z-score Label Binarization 1–4→Low 5–9→High Cross-Validation K-Fold Splits 10 (stratified) LOSO Subjects 20 (CLARE) 21 (CL-Drive) Split Strategy Subject-wise

Table 6: Data processing and augmentation configuration. Configuration Value Adapter Training Protocol Loss Function Smooth L1 (β=0.1) Optimizer AdamW Learning Rate 3 × 10−4 Weight Decay 10−4 LR Schedule Linear warmup (5 epochs) + Cosine annealing + Plateau reduction Plateau Factor 0.5 Plateau Patience 5 epochs Early Stop Patience 5 epochs Minimum Epochs 20 Batch Size 512 Gradient Accumulation 4 steps Effective Batch Size 2048 Mixed Precision FP16

Table 4: Progressive fine-tuning hyperparameters for K-Fold and LOSO configurations. Auto-α denotes inverse frequency class weights. Configuration Value Model Architecture Adapter Dropout 0.1 ECG-FM Encoder Dim 768 ECG-FM Encoder Layers 12 Classifier Hidden Dim 256 Classifier Dropout 0.2 Optimization Optimizer AdamW Weight Decay 1e−4 Scheduler Cosine + Warmup Warmup Ratio 0.1 Early Stop Patience 10 epochs Metric (K-fold) AUC Metric (LOSO) Macro F1 Mixed Precision Enabled

Table 5: Model architecture and optimization configuration.

F.2

Implementation Details

AdamW Optimizer Configuration We use the AdamW optimizer with default momentum parameters (β1 = 0.9, β2 = 0.999) and weight decay of 1×10−4 applied to all parameters except bias terms and batch normalization parameters. Cosine Annealing Schedule The learning rate follows a cosine annealing schedule with a warmup period of 10% of total training steps. During warmup, the learning rate linearly increases from 0 to the target learning rate. After warmup, the learning rate decays following:    t − twarmup 1 π ηt = ηmin + (ηmax − ηmin ) 1 + cos 2 T − twarmup (39) where ηmin = 0, ηmax is the initial learning rate, t is the current epoch, twarmup is the warmup duration, and T is the total number of epochs.

Table 7: Adapter training protocol configuration.

Early Stopping We implement early stopping with a patience of 10 epochs. For K-fold evaluation, the model is selected based on validation AUC to maximize discrimination ability. For LOSO evaluation, we use macro F1 score as the selection metric since it is more robust when validation sets contain imbalanced single-subject data.

G

Per-Subject Performance Analysis

To understand how CogAdapt’s ProFine strategy affects individual subjects under leave-one-subject-out (LOSO) evaluation, we provide detailed per-subject analysis on the CLARE dataset. This analysis reveals the heterogeneity in subjectindependent generalization and identifies which subjects benefit most from progressive fine-tuning. Figure 7 presents per-subject macro F1-scores across all three ProFine scenarios (A: frozen encoder, B: partial unfreeze, C: full fine-tuning), while Figure 8 shows the corresponding accuracy metrics. Several notable patterns emerge from this granular analysis. Substantial inter-subject heterogeneity. Performance varies dramatically across subjects, with F1-scores in Scenario C ranging from 0.591 (Subject 20) to 0.813 (Subject

Stage A Mean: 0.527 Stage B Mean: 0.623 Stage C Mean: 0.704

1.0

Stage A (Adapter Only) Stage B (Top-2 Layers) Stage C (Full Fine-tuning)

Macro F1-Score

0.8 0.6 0.4 0.2 0.0

1

2

3

4

5

6

7

8

9

10

11

Subject ID

12

13

14

15

16

17

18

19

20

Figure 7: Per-subject macro F1-score across all training stages under LOSO cross-validation on the CLARE dataset

9)—a 22.2 percentage point gap. This substantial variance underscores the challenge of subject-independent cognitive load assessment and validates our focus on LOSO evaluation rather than random k-fold splitting. Consistent progressive improvement. Most subjects exhibit monotonic performance gains across scenarios A→B→C. Subject 9 exemplifies this trend, achieving F1scores of 0.548 (A), 0.699 (B), and 0.813 (C)—a cumulative improvement of 26.5 percentage points. Similarly, Subject 8 progresses from 0.561 (A) to 0.694 (B) to 0.811 (C). This consistency across diverse physiological profiles suggests that ProFine’s layer-wise adaptation strategy effectively transfers clinical ECG knowledge to cognitive load patterns. Identification of challenging subjects. A subset of subjects (5, 6, 10, 15, 20) demonstrate relatively lower performance across all scenarios, with Scenario C F1-scores below 0.65. Subject 6 achieves only 0.622 F1 even under full finetuning, compared to the dataset mean of 0.702. These subjects may exhibit atypical physiological responses to cognitive load or higher signal noise, representing realistic deployment challenges for subject-independent systems. Accuracy-F1 divergence reveals class imbalance. Subject 11 presents a striking divergence: 0.978 accuracy but only 0.561 F1 in Scenario A. This pattern—high accuracy with moderate F1—indicates severe class imbalance for this participant, where the model achieves high accuracy by predicting the majority class while failing to detect the minority class. The focal loss employed in our training objective partially addresses this, as evidenced by improved F1 balance in Scenarios B (0.565) and C (0.737), though the gap persists. Differential benefit from fine-tuning. The magnitude of improvement from A→C varies substantially: Subject 9 gains +26.5pp F1, while Subject 15 improves only +12.6pp. Subjects with initially poor Scenario A performance (below 0.50 F1) tend to benefit more from fine-tuning, suggesting that frozen features transfer poorly for these individuals but taskspecific adaptation can recover performance. Conversely, subjects with strong Scenario A baselines (e.g., Subject 12: 0.622 F1) show smaller absolute gains (+13.4pp), indicat-

ing their physiological signatures already align well with pretrained representations. Robustness of top performers. Subjects 8, 9, and 13 consistently rank in the top quartile across all scenarios, achieving Scenario C F1-scores of 0.811, 0.813, and 0.801 respectively. These subjects likely exhibit canonical cognitive load signatures that align with the foundation model’s learned ECG patterns, making them easier to classify in subjectindependent settings.

H

Comparative Performance Distribution Analysis

To provide comprehensive insight into CogAdapt’s performance relative to baseline methods, we present detailed distributional analyses of macro-F1 scores under leave-onesubject-out (LOSO) cross-validation. Figures 9 and 10 visualize the complete statistical distributions for the CLARE and CL-Drive datasets, respectively, enabling direct comparison of model robustness and consistency across individual subjects.

H.1

CLARE Dataset Performance Distribution

Figure 9 presents box plot comparisons of F1 scores across all 20 subjects in the CLARE dataset under LOSO evaluation. CogAdapt (Scenario C) achieves a mean F1 score of 0.704 with a median of 0.707, substantially outperforming both baseline methods. The Transformer baseline obtains a mean of 0.496 (median: 0.503), while ECG-LightCNN achieves 0.514 (median: 0.525). This represents absolute improvements of +20.8 percentage points over Transformer and +19.0 percentage points over ECG-LightCNN, demonstrating the substantial benefit of foundation model adaptation for cognitive load assessment. The interquartile range (IQR) for CogAdapt spans from Q1=0.632 to Q3=0.757, indicating relatively consistent performance across the majority of subjects. In contrast, both baselines exhibit tighter but lower-performing distributions, with ECG-LightCNN showing an IQR of [0.495, 0.535] and

Stage A Mean: 0.681 Stage B Mean: 0.698 Stage C Mean: 0.765

1.0

Stage A (Adapter Only) Stage B (Top-2 Layers) Stage C (Full Fine-tuning)

Accuracy

0.8 0.6 0.4 0.2 0.0

1

2

3

4

5

6

7

8

9

10

11

Subject ID

12

13

14

15

16

17

18

19

20

Figure 8: Per-subject accuracy across all training stages under LOSO cross-validation on the CLARE dataset.

Transformer spanning [0.473, 0.526]. The whiskers for CogAdapt extend from 0.591 to 0.813, revealing considerable subject-to-subject variability—a characteristic challenge of LOSO evaluation that our approach navigates more successfully than training-from-scratch alternatives.

Figure 9: Performance distribution comparison on CLARE dataset under LOSO cross-validation. CogAdapt (Scenario C) achieves mean F1=0.704, substantially outperforming Transformer (F1=0.496) and ECG-LightCNN (F1=0.514) baselines. Box plots show median (black line), mean (red diamond), interquartile range (colored box), and whiskers extending to 1.5×IQR.

Notably, CogAdapt’s distribution exhibits a slight positive skew, with the mean (0.704) closely aligned with the median (0.707), suggesting that the majority of subjects benefit substantially from foundation model transfer. The baseline methods, conversely, show more symmetric distributions centered at significantly lower performance levels, indicating fundamental limitations in learning generalizable cognitive load signatures from limited training data.

H.2

CL-Drive Dataset Performance Distribution

Figure 10 illustrates the performance distributions on the CLDrive dataset across 21 subjects. CogAdapt maintains strong performance with a mean F1 score of 0.704 and median of 0.707. The CL-Drive Transformer baseline achieves a mean of 0.541 (median: 0.560), representing a +16.3 percentage point improvement for our approach. The CL-Drive CNN baseline obtains a mean of 0.607 (median: 0.600), which CogAdapt surpasses by +9.7 percentage points. The distributional characteristics reveal important insights into cross-dataset generalization. CogAdapt’s IQR spans [0.630, 0.762], demonstrating robustness comparable to CLARE despite the different task domain (driving simulation vs. multitasking). The Transformer baseline exhibits greater variability with an IQR of [0.493, 0.600] and whiskers extending from 0.425 to 0.669, suggesting inconsistent subjectindependent generalization. The CNN baseline shows a wider distribution with IQR [0.527, 0.666] and whiskers spanning [0.437, 0.810], indicating that while some subjects are classified effectively, performance degrades substantially for others. The superior mean and tighter distribution of CogAdapt across both datasets underscore the effectiveness of our progressive fine-tuning strategy (Scenario C) combined with LeadBridge adaptation. While both baselines struggle with subject-independent generalization—evidenced by their lower central tendencies and in some cases high variance— CogAdapt consistently achieves higher F1 scores across the majority of test subjects, validating the hypothesis that pretrained ECG foundation models capture more generalizable physiological signatures than models trained from scratch on limited cognitive load data.

H.3

Cross-Dataset Robustness

Comparing Figures 9 and 10 reveals that CogAdapt maintains consistent performance characteristics across both datasets, with identical mean (0.704) and median (0.707) values despite differences in task paradigm, participant demographics, and recording conditions. This consistency suggests that the

Figure 10: Performance distribution comparison on CL-Drive dataset under LOSO cross-validation. CogAdapt (Scenario C) achieves mean F1=0.704, outperforming CL-Drive Transformer (F1=0.541) and CL-Drive CNN (F1=0.607) baselines. Consistent performance across datasets demonstrates robust cross-domain generalization.

features learned through progressive fine-tuning of ECG-FM generalize not only across individuals within a dataset but also across different cognitive load elicitation protocols. In contrast, both baseline methods exhibit dataset-specific performance variations, with ECG-LightCNN showing notably different distributions between CLARE and CL-Drive, indicating potential overfitting to dataset-specific artifacts. The box plot visualizations demonstrate that foundation model adaptation via CogAdapt provides substantial and statistically robust improvements over training-from-scratch approaches for subject-independent cognitive load assessment from wearable ECG signals. The distributional analyses complement our main results (Table 1) by revealing not only superior average performance but also more favorable variance characteristics, which are critical for reliable deployment in real-world human-computer interaction applications.

I

K-Fold Cross-Validation Analysis

To complement the leave-one-subject-out (LOSO) evaluation presented in the main paper, we provide detailed Kfold cross-validation results that illuminate the performance characteristics of our progressive fine-tuning strategy under within-subject data splitting. Unlike LOSO evaluation, which tests cross-subject generalization, K-fold validation allows some subject data to appear in both training and validation sets, providing an upper bound on achievable performance and revealing the intrinsic difficulty of the cognitive load classification task independent of subject transfer challenges.

I.1

Performance Across Training Scenarios

Figures 11 and 12 present comprehensive K-fold crossvalidation results for the CLARE and CL-Drive datasets respectively, tracking three key metrics (Macro F1-Score, Accuracy, and AUROC) across all 10 folds for each ProFine scenario.

CLARE Dataset Progression. On the CLARE dataset (Figure 11), we observe consistent monotonic improvement across scenarios A→B→C. Scenario A (frozen encoder) achieves a mean macro F1-score of 0.717, establishing a strong baseline that already validates the transferability of clinical ECG features to cognitive load assessment. Scenario B (top-layer adaptation) improves this to 0.736 (+1.9pp), while Scenario C (full fine-tuning) reaches 0.785 (+6.8pp over A). Notably, AUROC shows the most dramatic gains, progressing from 0.843 (A) to 0.857 (B) to 0.898 (C), indicating that full fine-tuning substantially improves class separation in the learned feature space. The fold-wise trajectories reveal interesting patterns: Folds 8-10 consistently achieve the highest performance across all scenarios, with Fold 8 reaching F1=0.830 in Scenario C—the best single-fold result on CLARE. This suggests that certain combinations of subject data facilitate more effective learning, possibly due to reduced inter-subject physiological variability or more balanced class distributions within those particular splits. CL-Drive Dataset Characteristics. The CL-Drive results (Figure 12) demonstrate notably higher baseline performance compared to CLARE, with Scenario A achieving mean F1=0.821 versus 0.717 on CLARE—a 10.4pp advantage. This gap persists across all scenarios: Scenario C reaches F1=0.860 on CL-Drive versus 0.785 on CLARE. We attribute this to the more controlled recording environment of the driving simulator, which reduces motion artifacts and maintains consistent sensor placement across trials, resulting in higher-quality ECG signals. Interestingly, CL-Drive exhibits greater fold-to-fold variability than CLARE, particularly in Scenario A where F1 scores range from 0.790 (Fold 2) to 0.845 (Fold 10)—a 5.5pp spread versus CLARE’s 2.7pp range. This heightened variance likely reflects the diverse driving scenarios (highway, urban, adverse weather) that create heterogeneous cognitive load patterns across different experimental blocks.

I.2

Scenario-Specific Convergence Patterns

A striking observation emerges when examining the relative improvements between scenarios. On CLARE, the A→B transition yields modest gains (+1.9pp F1), while B→C produces substantially larger improvements (+4.9pp F1). This asymmetric pattern suggests that shallow adaptation of top encoder layers captures only incremental task-specific refinements, whereas full fine-tuning enables more comprehensive feature reorganization that better aligns with cognitive load signatures. Conversely, CL-Drive shows more balanced incremental gains: A→B yields +1.0pp and B→C adds +2.9pp. The smaller B→C improvement on CL-Drive, despite using identical hyperparameters and architecture, indicates that the frozen ECG-FM features already align reasonably well with driving-related cognitive load patterns, requiring less aggressive adaptation than the multitasking paradigm in CLARE.

I.3

Stability and Variance Analysis

The standard deviations across folds (shown as dotted horizontal lines in the figures) provide insight into model stability.

0.90

Stage A ( =0.717) Stage B ( =0.736) Stage C ( =0.785)

0.85

Stage A ( =0.742) Stage B ( =0.762) Stage C ( =0.813)

Stage A ( =0.843) Stage B ( =0.857) Stage C ( =0.898)

0.975 0.950

0.85

0.75

0.80

AUROC

Accuracy

Macro F1-Score

0.925 0.80 0.900 0.875 0.850

0.75

0.825

0.70 0.70

1

2

3

4

5

6

7

Fold Number

8

9

10

0.800 1

2

3

4

5

6

7

Fold Number

8

9

10

1

2

3

4

5

6

7

Fold Number

8

9

10

Figure 11: CLARE dataset K-fold cross-validation performance across 10 folds. Line charts show macro F1-score (left), accuracy (center), and AUROC (right) for each fold and training stage. Markers indicate individual fold results: circles for Stage A (teal), squares for Stage B (light blue), triangles for Stage C (lavender). Dotted horizontal lines mark stage-wise means. Stage C achieves consistent improvements across all folds (mean F1 = 0.785, accuracy = 0.813, AUROC = 0.898), with universal gains over Stages A and B. The relatively smooth progression across folds indicates stable performance, though Fold 8 and Fold 10 show notably stronger Stage C results (F1 ¿ 0.82), suggesting some sensitivity to specific train-test partitions.

On CLARE, Scenario C exhibits slightly elevated variance (std=0.034 for F1) compared to Scenario A (std=0.010), reflecting the increased model capacity and sensitivity to training data composition when all 91M parameters are trainable. However, this variance increase is modest relative to the substantial performance gains, maintaining a favorable stabilityperformance trade-off. CL-Drive demonstrates comparable variance patterns but with uniformly higher absolute performance across all metrics. The AUROC metric shows particularly impressive consistency in Scenario C, with 8 out of 10 folds exceeding 0.945 and a peak performance of 0.966 on Fold 9—approaching near-perfect discrimination between high and low cognitive load states under within-subject evaluation conditions.

I.4

K-Fold versus LOSO Performance Gap

Comparing these K-fold results to the LOSO evaluation in Table 1 (main paper) quantifies the subject-independent generalization challenge. On CLARE, Scenario C achieves F1=0.785 under K-fold but drops to 0.626 under LOSO— a degradation of 15.9pp. CL-Drive shows a smaller gap: F1=0.860 (K-fold) versus 0.768 (LOSO), representing 9.2pp degradation. This differential highlights that CLARE participants exhibit greater inter-subject physiological variability, making cross-subject transfer more challenging despite similar within-subject classification accuracy. The K-fold results thus establish an empirical upper bound for cognitive load assessment performance when subjectspecific data is available during training, motivating future work on meta-learning or few-shot adaptation approaches that could narrow the K-fold/LOSO gap for practical deployment scenarios where limited per-user calibration data might be collected.

Stage A ( =0.821) Stage B ( =0.831) Stage C ( =0.860)

0.925

0.96

0.875

0.875

0.94

0.825

AUROC

0.900

0.850

0.850 0.825

0.800

0.800

0.775

0.775

0.750

0.750

Stage A ( =0.906) Stage B ( =0.919) Stage C ( =0.945)

0.98

0.900

Accuracy

Macro F1-Score

0.925

1.00

Stage A ( =0.823) Stage B ( =0.833) Stage C ( =0.862)

0.92 0.90 0.88 0.86 0.84

1

2

3

4

5

6

7

Fold Number

8

9

10

1

2

3

4

5

6

7

Fold Number

8

9

10

1

2

3

4

5

6

7

Fold Number

8

9

10

Figure 12: CL-Drive dataset K-fold cross-validation performance across 10 folds. Similar visualization to Figure 11, showing consistently high baseline performance (Stage A mean F1 = 0.821, AUROC = 0.906) that improves with progressive training. Stage C achieves nearceiling performance (mean F1 = 0.860, AUROC = 0.945), with several folds exceeding F1 = 0.88 and AUROC = 0.96. The tighter clustering of fold performances (smaller variance) compared to CLARE suggests more homogeneous data characteristics in the driving simulation task. Folds 6, 8, and 9 consistently achieve the highest performance across all stages, indicating these particular train-test splits may have more separable cognitive load patterns.