Physiological Information Reliability: Cross-Layer Adaptive Resource Allocation for Cardiovascular Sensing

ML4H 2026 - Findings Track

Machine Learning for Health (ML4H) 2026

Physiological Information Reliability: Cross-Layer Adaptive Resource Allocation for Cardiovascular Sensing Navaneeth Krishnan K and Janakiraman K and Harinisri V ∗

arXiv:2609.00435v1 [eess.SP] 31 Aug 2026

Abstract Cardiovascular sensing systems must preserve clinically useful information despite signal degradation, wireless packet loss, energy constraints, and edge-computation latency, yet signal quality, delivery reliability, energy, and latency are conventionally optimized as separate, layer-local objectives. We formalize Physiological Information Reliability (PIR), a joint state representation Xt = [Sm , Sw , Se , Sc ] that couples a multimodal ECG/PPG physiological information value (PIV) with wireless-channel, energy, and compute-placement state, and instantiate its controller as a LinUCB contextual bandit that adapts network-coding redundancy and edge/local compute placement online. In a controlled, multiseed (n=5) evaluation of 1,200 sequential monitoring windows per seed under a Gilbert–Elliott burst-erasure channel, PIRLinUCB reaches the lowest steady-state energy draw (0.089 ± 0.004 J/window) and the highest mean reward (0.205 ± 0.065) among five compared policies, while every policy meets a 150 ms latency deadline; a hand-tuned heuristic baseline is statistically indistinguishable from it on reward. We report the accuracy–energy– latency trade-off honestly, including where the bandit’s advantage is condition-dependent rather than universal, and discuss the limits of a synthetic-data, simulated-channel evaluation. Code and configurations are provided for reproducibility. Keywords: physiological monitoring, contextual bandits, network coding, edge computing, wireless health sensing

Data and Code Availability This study uses no real patient data. All physiological signals are ∗ N. Krishnan K. is with the Department of Electronics and

Communication Engineering, Saveetha School of Engineering ( e-mail: knavaneeth385@gmail.com) and J. K is with the Government Kilpauk Medical College and Hospital ( e-mail: kjanaki2408@gmail.com) and Harinisri V. is with the Department of Physics and Nanotechnology, SRM Institute of Science and Technology, Chennai, India (e-mail: harinisriv2003@gmail.com).

© .

produced by a documented, parameterized synthetic ECG/PPG generator (Section 4) that reproduces the graded and complete-modality degradation protocol used in prior multimodal fusion work (Kamalakannan and Kamalakannan, 2026); no public corpus offers independently controllable, graded degradation of both modalities simultaneously. Code, configuration files, and the full multiseed experiment logs/figures underlying every number reported here are included as anonymized supplementary material for review, and will be released at a de-anonymized public repository upon acceptance. Institutional Review Board (IRB) This work involves no human subjects and no real patient data; all evaluation is performed on synthetic waveforms generated in software. No IRB review was required.

1. Introduction Wearable cardiovascular sensing pipelines chain together sensing, wireless transmission, edge or ondevice inference, and clinical decision support. Reliability in such pipelines is conventionally treated as an end-to-end packet-delivery question: did the bits arrive? But not every window of ECG/PPG data carries equal clinical value: a packet spanning a clean, diagnostically usable beat and a packet spanning motion-artifact noise are not equally worth protecting, yet systems that allocate redundancy, compute, or energy by fixed or purely channel-aware rules treat them identically. Existing work instead optimizes signal quality (Orphanidou et al., 2015; Kamalakannan and Kamalakannan, 2026), transmission reliability (Ho et al., 2006; Kamalakannan and Velmurugan, 2026), or task/semantic relevance of transmitted content (Gündüz et al., 2023) largely in isolation from one another and from the energy and compute budget that a body-worn device actually has. We argue that communication reliability is not the same objective as physiological information reliability, and that a

Physiological Information Reliability

cross-layer controller should allocate scarce wireless, energy, and compute resources in proportion to the physiological information at stake, not only to channel conditions. We formalize this as PIR: a joint state Xt = [Sm (t), Sw (t), Se (t), Sc (t)] spanning medical (signalquality/fusion), wireless, energy, and compute substates (Figure 1), and instantiate its control policy as a contextual bandit that learns, online, which coding redundancy and compute placement to use per monitoring window. PIR bridges two previously separate systems – a multimodal ECG/PPG fusion and signalquality front end (Kamalakannan and Kamalakannan, 2026) and an adaptive random-linear-networkcoding transport layer (Kamalakannan and Velmurugan, 2026) – into one closed-loop, learned controller rather than a hand-tuned heuristic. Contributions. (1) A formal PIR state representation connecting a physiological information value (PIV) with wireless, energy, and compute conditions (Section 3). (2) A LinUCB contextual-bandit controller that selects network-coding redundancy and edge/local compute placement to maximize a PIVweighted decode utility net of energy, latency, and power penalties. (3) A controlled, multiseed empirical comparison against four non-learned baselines that reports where the learned controller helps, where it does not, and what a synthetic-data proxy evaluation can and cannot show (Sections 5–7).

ECG

PPG

CardioFusion SQI → PIV

Edge / local inference

Xt = [Sm , Sw , Se , Sc ] medical / wireless / energy / compute

APC-RLNC encode + GE channel

PIR Controller (LinUCB)

Action reward at =(Rt , mt )

Figure 1: PIR-Framework architecture. ECG/PPG signal quality and cross-modal agreement give the physiological information value (PIV), which joins wireless, energy, and compute state into Xt . A LinUCB controller selects coding redundancy R and compute placement m; APC-RLNC-style coding and the Gilbert–Elliott channel determine delivery, which is scored and fed back as reward.

and Velmurugan (2026) combine RLNC with adaptive peer clustering over Gilbert–Elliott burst-erasure channels (Gilbert, 1960; Elliott, 1963). Our channel and RLNC decode-probability model reuse this formulation. Adaptive, ML-based resource allocation. Contextual bandits give low-regret online decision policies with linear-model efficiency (Li et al., 2010; Agrawal and Goyal, 2013) and have been applied to sequential mobile-health decisions (Tewari and Murphy, 2017). We apply the same machinery to sensing/coding/placement decisions rather than behavioral interventions. Existing approaches typically optimize one or two of {signal quality, delivery, energy, latency} in isolation; PIR instead treats preservation of physiologically valuable information as an explicit, joint control objective over sensing quality, wireless state, energy, and compute placement.

2. Related Work Physiological signal quality. Signal-quality indices (SQI) assess whether a reliable heart rate can be recovered from a given ECG/PPG segment (Orphanidou et al., 2015); Kamalakannan and Kamalakannan (2026) extend this to multimodal fusion under six graded and complete-modality degradation regimes, the protocol our PIV estimator follows. Task-aware information preservation. A growing communications literature argues that systems should preserve decision-relevant meaning, not raw bits (Gündüz et al., 2023). PIR instantiates this principle for one concrete closed-loop physiologicalsensing pipeline rather than proposing a general semantic-communication scheme.

3. The PIR Framework Joint state. Xt = [Sm (t), Sw (t), Se (t), Sc (t)]. Sm carries per-modality SQI, modality-presence flags, and a scalar physiological information value PIVt ∈ [0, 1]:
PIVt = clip wq SQIcomb,t − wa δt , 0, 1 ,

(1) Wireless medical sensing. Random linear network coding (RLNC) gives algebraic resilience to where SQIcomb is a presence-weighted average of packet erasures (Ho et al., 2006); Kamalakannan per-modality SQI (a missing modality is excluded, 2

Physiological Information Reliability

windows/seed (200 windows × 6 degradation regimes: both-clean, ECG-degraded, PPG-degraded, bothdegraded, ECG-missing, PPG-missing), following Kamalakannan and Kamalakannan (2026)’s protocol. Graded degradation adds severity-scaled Gaussian noise, sinusoidal baseline wander, a band-limited muscle-artifact component, and randomly-timed motion bursts; missing-modality windows are replaced by a structureless noise floor.

not scored as 0) and δt is the normalized crossmodal heart-rate disagreement, gated to 0 unless both modalities are present (wq =0.75, wa =0.25), following the availability/quality distinction of Kamalakannan and Kamalakannan (2026). Sw (t) = [SNRt , per t ] evolves via a two-state Gilbert–Elliott chain (Gilbert, 1960; Elliott, 1963) with per-step transition probabilities pgb , pbg ; SNR is drawn stateconditionally and mapped √ to a BPSK/AWGN biterror rate BER = Q( 2 · SNRlin ) and then to a packet-erasure probability per = 1 − (1 − BER)L over an L-bit payload, so SNR and erasure risk cannot silently disagree. Se (t) = [Bt , Epkt,t ] is residual battery and an additive per-packet energy cost E = etx ntx + erx nrx + ecomp nops ; Sc (t) is on-device vs. edge-offload compute latency plus RLNC transmission latency.

Wireless. A two-state Gilbert–Elliott channel (pgb =0.03, pbg =0.08; SNR ∼ N (18, 1.52 ) dB good, N (7, 12 ) dB bad; 64-byte coded payload) drives packet erasure identically for every policy (common random numbers, for variance reduction).

Energy / compute. A BLE-class radio (etx =2 mJ, erx =1 mJ/packet) and a 64 MHz on-body MCU vs. 1.4 GHz edge target (+12 ms Action and reward. The controller chooses at = RTT) parameterize Se , Sc ; deadline 150 ms, power (Rt , mt ) – RLNC redundancy (generation size K=32 cap 100 mW. fixed) and local/edge placement – from a 6-arm set R ∈ {2, 8, 20} × m ∈ {local, edge}. Delivery uses Baselines. Fixed-Rate (K=32, R=8, local, static); the closed-form RLNC decode probability (Ho et al., Always-Local / Always-Edge (R=20, placement
PK+R 2006), Pdec (K, R, p) = k=K K+R (1−p)k pK+R−k . pinned); Heuristic-Rule (battery-threshold placek We define a PIR utility proxy UPIR = PIVt · Pdec – ment; smallest R whose closed-form decode probabilhow much physiologically valuable information is ex- ity clears 0.99 at the current erasure estimate – same pected to be successfully decoded – and not as an physical features as the bandit, hand-written threshindependently validated measure of clinical informa- olds instead of a learned policy); PIR-LinUCB (α=0.3, ridge λ=1.0, λE =4.0, λL =0.05, λP =0.02). tion agreement. The reward is rt = UPIR −λE Et −λL sp(Lt −Lmax )−λP sp(Pt −Pmax ), Protocol and metrics. Each policy runs as one (2) continuous sequential monitoring session per seed with softplus sp(·) penalties that are near-zero inside over the identical channel/physiological stream. We the 150 ms deadline Lmax and 100 mW power cap report steady-state metrics (second half of each session) as mean ± 95% CI (Student-t, 5 seeds) to Pmax and grow smoothly once violated. separate converged decision quality from the banController. PIR-LinUCB (Li et al., 2010) main- dit’s own exploration cost. Primary metrics: HRtains, per arm a, ridge posterior θ̂a = A−1 a ba and MAE (bpm), energy/window (J), deadline-success selects rate. Secondary: packet delivery ratio (PDR), a q threshold-based bradycardia/tachycardia surrogate ⊤ −1 at = arg max x⊤ (3) event-detection F1, and mean reward. t θ̂a + α xt Aa xt , a

over the causal context xt (built from the previously observed channel telemetry, not the about-to- 5. Results be-realized draw), updated via a rank-1 Sherman– All five policies achieve 100% deadline-success and Morrison rule after each observed reward. stay well under the power cap (Table 1), so the differences below are a genuine energy/accuracy trade, not a constraint being quietly violated. PIR-LinUCB 4. Experimental Setup attains the lowest steady-state power (11.2 mW, vs. Data. A physiologically-grounded synthetic 12.0 mW Fixed/Heuristic and 15.0 mW AlwaysECG/PPG generator (125 Hz, 8 s windows) pro- Local/Edge) and the highest mean reward, narrowly duces n=5 independent seeds × 1,200 sequential ahead of Heuristic-Rule (+0.5%) and clearly ahead 3

Physiological Information Reliability

Table 1: Steady-state results, mean ± 95% CI over 5 seeds. Best per column in bold.

Policy Fixed-Rate Always-Local Always-Edge Heuristic-Rule PIR-LinUCB

HR-MAE (bpm)↓

Energy (J)↓

Reward↑

6.96 ± 1.11 5.95 ± 0.58 5.95 ± 0.58 6.08 ± 0.57 6.82 ± 1.62

0.096 ± 0.000 0.120 ± 0.000 0.120 ± 0.000 0.096 ± 0.006 0.089 ± 0.004

0.172 ± 0.054 0.113 ± 0.031 0.115 ± 0.031 0.204 ± 0.054 0.205 ± 0.065

Figure 3: Energy–latency trade-off (marker size ∝ PDR; stars = Pareto-optimal). PIRLinUCB and Heuristic-Rule dominate on energy and latency simultaneously; Always-Local approaches the 150 ms deadline.

6. Findings F1 – PIV is a usable cross-layer signal. PIV enters the controller’s context directly, and perregime results (Appendix B) show reward tracking PIV across all six regimes for every policy, consistent with PIV carrying decision-relevant information about downstream task difficulty. F2 – Adaptive allocation lowers energy without violating latency. PIR-LinUCB’s steady-state energy and power are the lowest of all five policies of the non-adaptive fixed policies (+19% vs. Fixed- while still meeting the 150 ms deadline on 100% of Rate, +81% vs. Always-Local/Edge). It does not windows. F3 – The advantage is condition-dependent, have the best raw HR-MAE: spending redundancy unconditionally (R=20) buys better raw accuracy at not universal. PIR-LinUCB’s reward edge over 34% more energy – the intended behavior of a reward Heuristic-Rule is within noise, and its raw accuracy that trades some accuracy for energy, not a deficiency. trails the redundancy-maximizing baselines. Here, Figures 2–3 show PIR-LinUCB and Heuristic-Rule learning recovers heuristic-competitive decisions with both lying on the accuracy–power and energy–latency less hand-tuning, not a dominant improvement over Pareto frontiers, with Fixed-Rate dominated on both a well-designed rule. axes. The reward gap between PIR-LinUCB and F4 – Static evaluation understates the case Heuristic-Rule (0.205 vs. 0.204) is far smaller than for prediction. The controller reacts only to the either policy’s own 95% CI half-width; with 5 seeds last-observed channel state (Section 3); a bursty this comparison would not survive a multiplicity- Gilbert–Elliott channel has short-horizon structure corrected significance test, and we report it as an a reactive bandit cannot exploit, motivating modeleffect size rather than a significance claim. predictive or forecasting extensions. Figure 2: Accuracy–power Pareto frontier (steadystate, pooled across seeds). PIR-LinUCB and Heuristic-Rule sit on the frontier at low power; Always-Local/Edge sit on it at the high-accuracy/high-power end; Fixed-Rate is dominated on both axes.

4

Physiological Information Reliability

7. Limitations

A random linear network coding approach to multicast. IEEE Transactions on Information Theory, 52(10):4413–4430, 2006.

Evaluation uses a controlled synthetic ECG/PPG generator, not recorded patient data; the front-end SQI/degradation protocol follows Kamalakannan and Kamalakannan (2026), whose signal-processing components are separately validated against real ICU and fetal-ECG recordings, but our own bandit-facing pipeline is not. The wireless, energy, and compute constants are documented and physically motivated but not fit to a specific real radio or MCU. UPIR is a PIR utility proxy (PIV × decode probability), not an independently validated measure of clinical information agreement, and the reported event-F1 uses a simple bradycardia/tachycardia threshold rather than a clinically adjudicated arrhythmia label. With 5 seeds, several pairwise differences (notably PIRLinUCB vs. Heuristic-Rule) do not clear a strict significance bar; we report effect sizes with honest CIs. We did not evaluate an oracle (perfect future-channelknowledge) upper bound in this submission; adding one, along with a state-ablation study (dropping PIV, wireless, energy, or compute context in turn) and real-channel/hardware validation, is the immediate next step. These results are an initial computational demonstration of physiological-information-aware resource allocation, not a clinical efficacy claim.

Navaneetha Krishnan Kamalakannan and Janakiraman Kamalakannan. CardioFusion-AI: Robust ECG–PPG fusion for multimodal physiological monitoring under signal degradation. arXiv preprint arXiv:2608.26000, 2026. Navaneetha Krishnan Kamalakannan and Harinisri Velmurugan. Adaptive peer clustering with hierarchical random linear network coding for resilient decentralized wireless networks. arXiv preprint arXiv:2608.26040, 2026. Lihong Li, Wei Chu, John Langford, and Robert E. Schapire. A contextual-bandit approach to personalized news article recommendation. In Proceedings of the 19th International Conference on World Wide Web (WWW), pages 661–670. ACM, 2010. Christina Orphanidou, Timothy Bonnici, Peter Charlton, David Clifton, David Vallance, and Lionel Tarassenko. Signal-quality indices for the electrocardiogram and photoplethysmogram: Derivation and applications to wireless monitoring. IEEE Journal of Biomedical and Health Informatics, 19 (3):832–838, 2015. Ambuj Tewari and Susan A. Murphy. From ads to interventions: Contextual bandits in mobile health. In James M. Rehg, Susan A. Murphy, and Santosh Kumar, editors, Mobile Health: Sensors, Analytic Methods, and Applications, pages 495–517. Springer, 2017.

References Shipra Agrawal and Navin Goyal. Thompson sampling for contextual bandits with linear payoffs. In Proceedings of the 30th International Conference on Machine Learning (ICML), volume 28 of Proceedings of Machine Learning Research, pages 127– 135, 2013.

Appendix A. Extended Configuration

E. O. Elliott. Estimates of error rates for codes on burst-noise channels. Bell System Technical Journal, 42(5):1977–1997, 1963.

LinUCB: α=0.3, ridge λ=1, context dimension 10 (per-modality SQI/presence, PIV, normalized SNR, EWMA erasure estimate, link state, battery fraction, E. N. Gilbert. Capacity of a burst-noise channel. Bell bias). Reward weights: λE =4.0 J−1 , λL =0.05 ms−1 , System Technical Journal, 39(5):1253–1265, 1960. λP =0.02 mW−1 , softplus β=0.5. Compute model: 8 GF(256) cycles/op, 2×106 (6.4× that for edge) inferDeniz Gündüz, Zhijin Qin, Inaki Estella Aguerri, ence cycles, 250 kbps PHY rate. All parameters and Harpreet S. Dhillon, Zhaohui Yang, Aylin Yener, their physical justification are documented in the reKai Kit Wong, and Chan-Byoung Chae. Beyond leased configuration file. transmitting bits: Context, semantics, and taskoriented communications. IEEE Journal on Selected Areas in Communications, 41(1):5–41, 2023. Appendix B. Per-Regime Breakdown Tracey Ho, Muriel Médard, Ralf Koetter, David R. Steady-state reward is highest in both-clean and Karger, Michelle Effros, Jun Shi, and Ben Leong. missing-modality regimes (PIR-LinUCB: 0.268, 5

Physiological Information Reliability

0.265, 0.258 for both-clean/ECG-missing/PPGmissing) and lowest under both-degraded (0.089), tracking mean PIV (0.739 both-clean vs. 0.543 both-degraded) across every policy, not only the bandit – consistent with F1. HR-MAE follows the same ordering (3.5–3.8 bpm clean/missing vs. 14.4 bpm both-degraded), since graded noise, not modality absence, is the harder regime for heart-rate estimation in this synthetic protocol.

6

Related documents

• Physiological Information Reliability: Cross-Layer Adaptive Resource Allocation for Cardiovascular Sensing#819883
• Physiological Information Reliability: Cross-Layer Adaptive Resource Allocation for Cardiovascular Sensing#809887
• Complexity of the Airflow in COPD#830120
• Site-Aware federated learning via embedding and resampling with electrocardiograms#359173
• Emerging Evidence on Molecular Dysregulation Driving Cardiomyopathies and Heart Failure#1068013