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

• Conventional reliability treats delivery as whether bits arrive, although ECG/PPG windows differ in clinical value.Source: §1 · p. 1
• Prior work largely separates signal quality, transmission reliability, semantic relevance, energy, and computation.Source: §1 · p. 1
• PIR allocates wireless, energy, and compute resources according to physiological information at stake rather than channel conditions alone.Source: §1 · p. 2
• PIR represents medical, wireless, energy, and compute substates jointly as Xt and uses a contextual bandit to select coding redundancy and compute placement.Source: §1 · p. 2
• The study compares PIR-LinUCB with four non-learned baselines using controlled multiseed experiments and reports both benefits and limits of synthetic proxy evaluation.Source: §1 · p. 2

Physiological signal quality

• Multimodal signal-quality work evaluates ECG/PPG reliability across graded and complete-modality degradation regimes.Source: §2 · p. 2

Task-aware information preservation

• Task-aware communications prioritizes decision-relevant meaning over raw bits, whereas PIR applies this principle to physiological sensing.Source: §2 · p. 2

Wireless medical sensing

• RLNC provides algebraic resilience to packet erasures, and adaptive peer clustering extends it to Gilbert–Elliott burst-erasure channels.Source: §2 · p. 2

Adaptive, ML-based resource allocation

• The framework combines ECG/PPG information estimation, adaptive network coding, and contextual-bandit decisions over sensing, coding, and placement.Source: Figure 1 · p. 2, §2 · p. 2

PIR's distinction

• PIR treats physiologically valuable information as an explicit joint objective over sensing quality, wireless state, energy, and compute placement.Source: §2 · p. 2

Joint state

• The medical substate contains per-modality SQI, modality-presence flags, and a scalar PIV, with missing modalities excluded from the quality average.Source: §3 · p. 2
• PIV combines presence-weighted signal quality with cross-modal heart-rate disagreement, which is considered only when both modalities are present.Source: §3 · pp. 2–3
• Wireless state follows a Gilbert–Elliott channel, while energy tracks battery and packet costs and compute state includes local, edge, and transmission latency.Source: §3 · p. 3

Action and reward

• The controller chooses among six actions combining RLNC redundancy R ∈ {2, 8, 20} with local or edge placement.Source: §3 · p. 3
• The utility proxy UPIR = PIVt · Pdec estimates physiologically valuable information successfully decoded, but is not independently validated clinical-information agreement.Source: §3 · p. 3
• Softplus penalties remain near zero within the 150 ms latency deadline and 100 mW power cap, then increase after violations.Source: §3 · p. 3

Controller

• PIR-LinUCB maintains a ridge posterior for each action and updates it from observed contextual rewards.Source: §3 · p. 3

Data

• Each seed contains 1,200 sequential 8-second windows across six ECG/PPG degradation regimes generated at 125 Hz.Source: §4 · p. 3
• Graded degradation adds scaled noise, baseline wander, muscle artifacts, and motion bursts, while missing modalities become structureless noise floors.Source: §4 · p. 3

Wireless

• All policies share a two-state Gilbert–Elliott channel with common random numbers, ensuring identical wireless randomness across comparisons.Source: §4 · p. 3

Energy / compute

• The setup models a BLE-class radio, on-body MCU, edge target with +12 ms round-trip time, a 150 ms deadline, and a 100 mW power cap.Source: §4 · p. 3

Baselines

• Baselines include Fixed-Rate, Always-Local, Always-Edge, and Heuristic-Rule alongside PIR-LinUCB.Source: §4 · p. 3

Protocol and metrics

• Policies run continuous sessions per seed, with steady-state results reported as means and 95% confidence intervals across five seeds.Source: §4 · p. 3

• 100% deadline-success was achieved by all five policies, with PIR-LinUCB using the lowest steady-state power at 11.2 mW.Fixed/Heuristic used 12.0 mW, while Always-Local/Edge used 15.0 mW.Source: §5 · pp. 3–4
• PIR-LinUCB and Heuristic-Rule lie on the accuracy–power and energy–latency Pareto frontiers, while Fixed-Rate is dominated on both axes.Source: §5 · p. 4
• R=20 redundancy improves raw accuracy but consumes 34% more energy, illustrating the intended accuracy–energy trade-off.Source: §5 · p. 4
• PIR-LinUCB’s mean reward is 0.205 versus 0.204 for Heuristic-Rule, a gap smaller than either policy’s 95% CI half-width.The comparison is reported as an effect size rather than a significance claim.Source: §5 · p. 4

• PIV tracks reward across all six degradation regimes, consistent with decision-relevant information about downstream task difficulty.Source: §6 · p. 4
• PIR-LinUCB has the lowest steady-state energy and power of all five policies while meeting the 150 ms deadline on 100% of windows.Source: §6 · p. 4
• PIR-LinUCB’s reward edge over Heuristic-Rule is within noise, and its raw accuracy trails redundancy-maximizing baselines.The reported advantage is heuristic-competitive with less hand-tuning, not dominant.Source: §6 · p. 4
• Because the controller uses only the last-observed channel state, it cannot exploit short-horizon structure in the bursty Gilbert–Elliott channel.This motivates model-predictive or forecasting extensions.Source: §6 · p. 4

Scope and validation

• The bandit-facing pipeline is evaluated on synthetic ECG/PPG waveforms rather than recorded patient data and is not itself validated against real recordings.The underlying signal-processing components were separately validated against real ICU and fetal-ECG recordings, but that validation does not extend to this pipeline.Source: §7 · p. 5

Proxy measures

• UPIR is a proxy defined as PIV multiplied by decode probability, not an independently validated measure of clinical information agreement.The event-F1 metric also uses a bradycardia/tachycardia threshold rather than clinically adjudicated arrhythmia labels.Source: §7 · p. 5

System modeling

• Wireless, energy, and compute constants are physically motivated and documented but are not fitted to a specific real radio or MCU.This limits direct interpretation of the simulated communication and hardware dynamics.Source: §7 · p. 5

Statistical power

• With 5 seeds, several pairwise differences, notably PIR-LinUCB versus Heuristic-Rule, do not clear a strict significance threshold.The study reports effect sizes with confidence intervals rather than significance claims for these comparisons.Source: §7 · p. 5

Unresolved validation

• The submission does not evaluate an oracle upper bound or state ablations, and it leaves real-channel and hardware validation for future work.The proposed next steps include dropping PIV, wireless, energy, or compute context in turn and testing on real channels and hardware.Source: §7 · p. 5

Bandit configuration

• LinUCB uses α=0.3, ridge λ=1, and a 10-dimensional context.The context includes per-modality SQI/presence, PIV, normalized SNR, EWMA erasure estimate, link state, battery fraction, and bias.Source: Appendix A · p. 5

Reward configuration

• Reward weights are λE=4.0 J−1, λL=0.05 ms−1, and λP=0.02 mW−1, with softplus β=0.5.These weights define the relative energy, latency, and power terms in the reward configuration.Source: Appendix A · p. 5

System model

• The compute model assumes 8 GF(256) cycles/op, 2×10^6 inference cycles, 6.4× that for edge execution, and a 250 kbps PHY rate.All parameters and their physical justification are documented in the released configuration file.Source: Appendix A · p. 5

PIV by regime

• 0.739 mean PIV in both-clean conditions falls to 0.543 in both-degraded conditions across policies.The reward ordering tracks this PIV ordering across every policy, not only the bandit.Source: Appendix B · p. 6

Heart-rate estimation

• 3.5–3.8 bpm HR-MAE in clean and missing-modality regimes rises to 14.4 bpm under both-degraded conditions.In this synthetic protocol, graded noise is harder for heart-rate estimation than modality absence.Source: Appendix B · p. 6