PaperlayerView exact revision on arXiv

Source-linked AI summary

CardioFusion-AI: Robust ECG--PPG Fusion for Multimodal Physiological Monitoring Under Signal Degradation

Navaneetha Krishnan Kamalakannan, Janakiraman Kamalakannan

arXiv:2608.26000v1eess.SPcs.LGq-bio.QM

TL;DR

ECG–PPG fusion mechanisms have been insufficiently compared under controlled, graded degradation and complete modality loss. CardioFusion-AI evaluates eight matched strategies across these conditions and finds attention fusion descriptively strongest overall, while adaptive gates detect absent modalities but poorly track graded signal quality.

  • Problem

    Comparative evidence on whether adaptive ECG–PPG fusion tracks graded degradation, rather than only modality absence, remains limited under matched controlled evaluation.

  • Method

    CardioFusion-AI combines a validated ECG–PPG signal-processing front end with a matched-protocol comparison of eight fusion architectures across six synthetic degradation regimes.

  • Results

    Attention fusion achieved the lowest descriptive overall MAE at 1.66 ± 0.43 bpm, but its advantage was not uniform across degradation types.

  • Takeaways & Limitations

    Modality availability and modality quality are distinct problems: adaptive gates detect complete loss but show little evidence of tracking graded quality.

  • Takeaways & Limitations

    Five training seeds support directional trends and effect sizes but not confirmatory pairwise significance after multiplicity correction.

Abstract

from arXiv · show

Wearable electrocardiogram (ECG) and photoplethysmogram (PPG) sensors are complementary but individually fragile: motion artifact, poor contact, and sensor dropout can degrade one or both signals. Fusion strategies that assume both modalities are equally trustworthy can become less reliable than a single clean modality under degradation. We present CardioFusion-AI, a framework whose signal-processing front end, including R-peak and systolic-peak detection, an Orphanidou-type signal-quality index, and beat-by-beat pulse transit time estimation, is validated on 53 real intensive-care recordings (848 windows; heart-rate mean absolute error 1.61 bpm for ECG and 2.78 bpm for PPG) and a real annotated fetal ECG database (R-peak F1 0.89-0.98). We then conduct a controlled synthetic degradation study comparing eight ECG-PPG fusion strategies across six degradation regimes spanning graded corruption and complete modality loss, using five independent training seeds. Attention fusion achieved the lowest descriptive overall error (1.66+/-0.43 bpm). Both adaptive gates reallocated weight toward the healthy modality under complete modality loss, but showed near-zero correlation between gate weight and signal quality under graded degradation (r = 0.10-0.24). Signal-quality conditioning produced a specific improvement under missing-PPG conditions (1.56+/-0.59 bpm), approaching the 1.48 bpm unimodal ceiling. With only five training seeds, no pairwise comparison survives Holm-corrected significance testing; effect sizes and confidence intervals are therefore reported. These results indicate that modality availability and modality quality are functionally distinct problems for adaptive fusion.

I. INTRODUCTION

CardioFusion-AI addresses the lack of systematic, statistically accountable comparisons of ECG–PPG fusion under controlled, graded degradation and modality loss. It combines a signal-quality-validated processing front end with matched evaluation of eight fusion strategies across six degradation regimes.

  • Motivation: ECG and PPG provide complementary cardiac information, but motion artifact, poor contact, perspiration, and sensor dropout can degrade either modality, sometimes simultaneously.ECG supports direct cardiac timing, while PPG provides pulse-waveform information for heart rate, pulse timing, and PTT estimation.
  • Motivation: Fusion that assumes equal modality trustworthiness can become less reliable than a single clean modality when one input deteriorates.This motivates adjusting each modality’s contribution according to its current signal quality.
  • Research gap: Existing evaluations rarely systematically test whether adaptive, signal-quality-conditioned gates track graded degradation rather than only modality presence or absence.The paper identifies a need for controlled protocols with matched architectures, matched training procedures, and explicit statistical-power accounting.
  • Study design: CardioFusion-AI combines a validated front end—peak detection, an Orphanidou-type SQI, and beat-by-beat PTT estimation—with a controlled synthetic degradation study.The study compares eight ECG–PPG fusion architectures sharing an identical encoder backbone and training protocol across six degradation regimes.
  • Study objective: The study tests whether signal-quality-aware fusion improves robustness under ECG and/or PPG degradation, while recognizing that explicit quality conditioning may provide only partial additional support.Its protocol uses physiologically grounded synthetic signals with independently controllable, graded per-modality corruption.

II. METHODS · A. Overview and Two-Part Evaluation Strategy

CardioFusion-AI uses two epistemically distinct evidence streams kept separate throughout the paper: real-recording validation for the signal-processing front end and controlled synthetic evaluation for fusion robustness under degradation.

  • A. Overview and Two-Part Evaluation Strategy: Two epistemically distinct evidence categories are kept explicitly separate throughout CardioFusion-AI.The separation structures the paper’s evaluation strategy.
  • A. Overview and Two-Part Evaluation Strategy: Claim A evaluates the signal-processing front end against real, third-party physiological recordings.This claim is supported in Section II-B.
  • A. Overview and Two-Part Evaluation Strategy: The signal-processing front end includes peak detection, signal-quality-index computation, and pulse-transit-time estimation.These components define Claim A.
  • A. Overview and Two-Part Evaluation Strategy: Claim B compares fusion architectures’ robustness under degradation.This claim concerns comparative robustness rather than front-end validation.
  • A. Overview and Two-Part Evaluation Strategy: Fusion robustness is studied with a controlled synthetic protocol across Sections II-C–II-G.The protocol supplies the evaluation setting for Claim B.
  • A. Overview and Two-Part Evaluation Strategy: The synthetic protocol is used because no public dataset provides graded, independently controllable, ground-truth-labeled degradation of both ECG modalities.This limitation motivates the controlled design for Claim B.

B. Real-Data Validation of the Signal-Processing Pipeline · C. Controlled Synthetic Degradation Study: Signal Generation

The signal-processing front end was validated on independent real datasets, while synthetic ECG–PPG windows enabled controlled degradation experiments with exact heart-rate ground truth and leakage-free splits.

  • B. Real-Data Validation of the Signal-Processing Pipeline: 0.89–0.98 F1 was achieved for direct-lead R-peak detection against cardiologist-verified fetal ECG annotations.Applying the detector naively to noisier abdominal leads reduced F1 to 0.19–0.34; per-record F1 correlated with SQI pass rate at Pearson r = 0.927.
  • B. Real-Data Validation of the Signal-Processing Pipeline: 1.61 bpm MAE was obtained for ECG-derived heart rate, compared with 2.78 bpm MAE for PPG-derived pulse rate, on 848 ICU windows.The corresponding monitor-reference correlations were 0.911 for ECG and 0.879 for PPG across 53 patients.
  • B. Real-Data Validation of the Signal-Processing Pipeline: 113.9 ms mean PTT was estimated beat by beat across 7,868 individually matched beats, with a median of 104.0 ms and SD of 55.6 ms.The estimator was validated on the same 53-subject BIDMC cohort after correcting a boundary-artifact bug in an earlier whole-window method.
  • B. Real-Data Validation of the Signal-Processing Pipeline: The front end and SQI computation were grounded in independently verifiable real-data performance before being used in the synthetic fusion-degradation comparison.The SQI served as an input to the signal-quality-conditioned fusion gate.
  • C. Controlled Synthetic Degradation Study: Signal Generation: Controlled synthetic data were generated because public datasets lack simultaneous ECG–PPG recordings with independently controllable, graded degradation and known physiological ground truth.Real degraded recordings confound degradation type, severity, and physiological state, preventing controlled comparison of fusion mechanisms.
  • C. Controlled Synthetic Degradation Study: Signal Generation: NeuroKit2 generated the physiologically grounded signals using the McSharry ECG dynamical model and PPG parameters for motion artifact, baseline drift, and burst noise.These components supplied the controlled corruption mechanisms used for the synthetic study.
  • C. Controlled Synthetic Degradation Study: Signal Generation: 8 s windows at 125 Hz contained clean ECG and PPG waveforms sharing a target heart rate drawn from HR ∼U(50, 110) bpm.Each modality was independently assigned clean, mild, moderate, severe, or missing severity; missing denoted complete dropout modeled as a noise floor without cardiac structure.
  • C. Controlled Synthetic Degradation Study: Signal Generation: 792 training, 234 validation, and 324 test window pairs were generated as i.i.d. draws with disjoint random-seed ranges across splits.The design prevented subject-level leakage because windows had no subject identity and supported the eight-architecture × five-seed sweep.

D. Signal Quality Assessment Applied to Synthetic Windows

The synthetic-window assessment used an Orphanidou-type SQI as both a diagnostic and an input to one adaptive fusion gate, while treating signal quality as a waveform-derived descriptor rather than ground truth. Mean SQI declined with graded degradation and collapsed under complete dropout, although the index can fail to discriminate in some conditions.

  • SQI definition and use: The Orphanidou-type SQI combined beat detection, template-correlation scoring, and feasibility checks for physiologically plausible rate and beat-to-beat regularity.It was used diagnostically and as an input feature to the signal-quality-conditioned fusion gate.
  • Shared architecture: Each modality was independently preprocessed, quality-assessed, and encoded before one of eight fusion mechanisms combined their embeddings into a physiological estimate.The SQI was an explicit gate input only for the SQI-conditioned adaptive variant.
  • SQI behavior under degradation: 0.988 to 0.946 for ECG and 0.995 to 0.952 for PPG as degradation increased from level 0 to level 3.Mean SQI decreased monotonically with degradation severity for both modalities, but the SQI is a waveform-derived descriptor rather than a ground-truth quality label and can fail to discriminate in some conditions.
  • SQI behavior under dropout: 0.19 mean SQI and 27% feasibility-pass rate occurred with missing ECG, versus 0.05 mean SQI and 14% feasibility-pass rate with missing PPG.These sharp collapses under complete dropout confirmed that the SQI was directionally responsive to modality loss.

E. Fusion Architectures

All eight fusion strategies use the same dual-stream, single-modality encoder backbone, isolating the fusion mechanism as the only varying factor. The architectures span unimodal, fixed, feature-level, global-weighted late, attention, and implicit adaptive-gate designs.

  • Shared backbone: The shared dual-stream architecture differs only in the fusion mechanism g(·) that combines the two per-modality embeddings.Holding the encoder architecture fixed isolates fusion-mechanism effects.
  • Shared backbone: All eight architectures share a four-block 1-D CNN encoder with 16–32–64–64 channels, dropout 0.2, global average pooling, and d = 64 embeddings.The encoder includes kernel size 7, batch normalization, ReLU, and max-pooling after each block.
  • Fusion strategies: The comparison includes ECG-only and PPG-only prediction, fixed-average fusion, feature-level fusion, and global-weighted late fusion.Fixed-average fusion uses a static, input-independent 50/50 combination, while global-weighted late fusion uses one learned weight pair shared across samples.
  • Fusion strategies: Attention fusion uses four-head bidirectional cross-attention between pooled ECG and PPG embeddings, with input-dependent attention weights that are not explicitly quality-supervised.The attended embeddings are concatenated for regression.
  • Fusion strategies: The implicit adaptive gate produces per-sample modality weights from the concatenated embeddings alone.Its gate is defined as g = softmax(MLP([e_ecg; e_ppg])).

F. Experimental Protocol · G. Statistical Analysis

The experimental protocol trains all eight architectures under five independent seeds with fixed data partitions. Statistical analysis reports seed-level test MAE summaries and uses pre-registered architecture comparisons across overall and regime-specific metrics.

  • F. Experimental Protocol: Five random seeds independently trained each of the eight architectures, while train/validation/test partitions were fixed and shared across models.Seeds controlled parameter initialization and minibatch ordering; splits were generated once using disjoint random-number-generator seed ranges.
  • F. Experimental Protocol: Adam optimization used a 5×10^-4 learning rate, 10^-5 weight decay, and gradient-norm clipping at 1.0.These settings were introduced after an initial pilot run.
  • G. Statistical Analysis: For each architecture and degradation regime, test MAE was summarized by its mean and standard deviation across five seeds.The seed was defined as one independently initialized and trained model instance.
  • G. Statistical Analysis: Three pre-registered architecture comparisons were evaluated across overall MAE and six regime-specific MAEs, totaling 21 tests.The comparisons were SQI-conditioned versus implicit adaptive gate, SQI-conditioned adaptive gate versus attention, and implicit adaptive gate versus attention.
  • G. Statistical Analysis: Fig. 2 plots test MAE against graded degradation severity from 0, clean, to 3, severe, while holding the partner modality clean.Error bars show standard deviation across five seeds.
  • G. Statistical Analysis: Statistical quantities were computed per seed and then averaged to avoid pseudo-replication.This procedure addresses the same pseudo-replication problem identified in the statistical analysis.

III. RESULTS … C. Robustness to Complete Modality Loss

Across degradation regimes, attention fusion had the lowest descriptive overall MAE, while complete modality loss most clearly separated strategies. Adaptive SQI conditioning approached the unimodal ceiling when PPG was missing but remained less robust when ECG was missing.

  • A. Overall Fusion Performance: 1.66 ± 0.43 bpm was attention fusion’s lowest descriptive overall MAE across six degradation regimes and five training seeds.Implicit and SQI-conditioned adaptive gates followed at 2.06 ± 0.51 bpm and 1.99 ± 0.61 bpm, respectively.
  • A. Overall Fusion Performance: 2.18–2.87 bpm was the descriptive MAE range for fixed-weight fusion variants and single-modality baselines, worse than attention fusion.These comparisons are reported descriptively rather than as inferentially significant differences.
  • B. Performance Across Degradation Severity: Under graded PPG degradation, attention fusion remained strongest at every severity level, whereas its advantage under graded ECG degradation eroded and reversed with increasing severity.Performance trajectories differed by degradation type and fusion mechanism rather than degrading uniformly.
  • C. Robustness to Complete Modality Loss: Complete modality loss separated architectures far more sharply than graded degradation.This pattern is shown in the rightmost two groups of the degradation-regime results.
  • C. Robustness to Complete Modality Loss: 14.34 bpm and 15.02 bpm were the catastrophic errors when ECG-only and PPG-only baselines lost their own modalities, respectively.The same baselines reached near-ceiling errors of 1.48 bpm with PPG missing for ECG-only and 1.47 bpm with ECG missing for PPG-only.
  • C. Robustness to Complete Modality Loss: 5.77–7.28 bpm was the fixed-weight fusion error range under either missing-modality condition because dead-channel contributions could not be reduced at inference.These strategies therefore could not approach the surviving-modality ceiling.
  • C. Robustness to Complete Modality Loss: 1.56 ± 0.59 bpm was the SQI-conditioned adaptive gate’s error with missing PPG, within 0.08 bpm of the 1.48 bpm unimodal ceiling.With missing ECG, the same gate reached 4.36 ± 1.59 bpm, better than the implicit gate’s 5.38 ± 4.11 bpm.

D. Adaptive Gate Behavior and Effect of Explicit SQI Conditioning

The adaptive gates sharply reallocated weight only when one modality was completely lost, while retaining a strong pooled association between ECG gate weight and relative signal quality. Explicit SQI conditioning produced higher ECG allocation when both modalities were available, but the supplied passages do not report its complete graded-degradation analysis.

  • Adaptive gate behavior: Both gates allocated most weight to ECG when both modalities were present, with implicit weights of 0.74–0.77 and SQI-conditioned weights of 0.975–0.998.These ranges spanned clean and degraded-but-present regimes.
  • Adaptive gate behavior: Both gates sharply reallocated weight only under complete modality loss, assigning >98% weight to the surviving modality.For the implicit gate, ECG weights were 0.009 with ECG missing and 1.000 with PPG missing.
  • Signal-quality association: 0.68 and 0.70 were the pooled mean Pearson correlations between ECG gate weight and SQI difference for the implicit and SQI-conditioned gates, respectively.The calculation included all regimes, including missing-modality extremes, and was performed per seed across test windows.

E. Statistical Significance

Across 21 Holm-corrected seed-level comparisons, no comparison survived correction. The results identify modality availability and quality as distinct adaptive-fusion problems while highlighting the limitations of a five-seed study.

  • Statistical significance: No comparison survived Holm correction across all 21 pre-registered seed-level paired tests (n = 5).This was treated as an explicit limitation of the five-seed study rather than a null result to minimize.
  • Statistical significance: p = 0.0625 was the minimum possible exact two-sided Wilcoxon value at n = 5 for the two nominally closest comparisons.These comparisons showed consistent direction across all five seeds.
  • Statistical significance: Cohen’s dz = 2.51 and 1.50 were large paired effect sizes for those two nominally closest comparisons.The comparisons involved the SQI-conditioned adaptive gate versus attention fusion under both-clean and PPG-degraded-only regimes.
  • Mechanistic finding: Modality availability and modality quality are functionally distinct problems for adaptive fusion.A validated quality descriptor as an input feature did not by itself guarantee proportional network use.

IV. DISCUSSION … C. Why SQI conditioning helped selectively

The discussion attributes attention fusion’s strongest descriptive performance to flexible bidirectional cross-attention and greater fusion capacity, while explaining why adaptive gates detected complete dropout more readily than graded degradation. SQI conditioning helped most under complete PPG dropout, where SQI changes were unambiguous, but did not materially alter graded-degradation weight allocation.

  • A. Why attention fusion performed best, with a specific exception: Attention fusion varies each modality’s contribution through bidirectional cross-attention without requiring an explicit quality signal.It retains both attended representations by concatenating them rather than compressing them into one convex combination.
  • A. Why attention fusion performed best, with a specific exception: Attention fusion’s greater representational capacity at the fusion step is consistent with its descriptively strongest overall performance.Unlike the gates, it retains both attended representations rather than reducing them to a single convex combination.
  • B. Why adaptive gating did not track graded quality: The gates learn from a 64-dimensional pooled embedding, optionally augmented with a 4-dimensional SQI vector, using a single scalar heart-rate loss.These architectures are trained end to end.
  • B. Why adaptive gating did not track graded quality: Complete modality dropout creates a large, easily separable embedding-space signal and an extreme SQI value, unlike the subtler shift from graded degradation.This asymmetry plausibly explains why adaptive gating responds more clearly to modality presence or absence than to graded quality.
  • C. Why SQI conditioning helped selectively: Explicit SQI conditioning produced its clearest benefit under complete PPG dropout, when PPG SQI collapsed to near zero.The gate can separate this regime easily because the SQI signal is least ambiguous there.
  • C. Why SQI conditioning helped selectively: Under graded degradation, SQI-conditioned and implicit gates showed no materially different weight-allocation behavior.Smaller SQI changes and a regression task that may remain solvable without a large weight shift help explain the selective benefit.

D. Modality competition as a plausible contributing factor · V. LIMITATIONS · VI. CONCLUSION

The study identifies modality competition as a plausible cause of poor PPG-specific learning despite correct gating during ECG dropout, while concluding that fusion robustness depends strongly on mechanism choice and modality conditions. The limited seed count and synthetic degradation protocol constrain confirmatory significance and real-world error interpretation.

  • D. Modality competition as a plausible contributing factor: ∼99% SQI-conditioned gate weight toward PPG during ECG dropout still yielded 4.36 bpm accuracy, versus 1.47 bpm for the dedicated PPG-only model.The gate’s decision was essentially correct, yet performance remained substantially worse than the identical-window unimodal baseline.
  • D. Modality competition as a plausible contributing factor: Because both gates favor ECG for most training windows, the PPG-specific encoder may receive systematically less gradient signal during joint training.This is presented as a plausible explanation for the gap between gate allocation and PPG-fusion accuracy.
  • V. LIMITATIONS: Five independent training seeds support consistent-direction trends and effect sizes but not confirmatory pairwise significance after multiplicity correction.The study therefore treats the observed comparisons as descriptive rather than definitive hypothesis tests.
  • V. LIMITATIONS: The two largest observed effects, dz > 1.5, are promising candidates for confirmation in a higher-replicate follow-up.The passage frames these effects as needing additional replication rather than as confirmed differences.
  • V. LIMITATIONS: The controlled, physiologically grounded degradation model approximates real-world corruption rather than reproducing it directly.Consequently, absolute error magnitudes may not transfer directly to unconstrained wearable conditions.
  • VI. CONCLUSION: Eight ECG–PPG fusion strategies were evaluated across graded corruption and complete modality loss using a front end independently validated against real ICU and fetal ECG recordings.The conclusion combines controlled robustness testing with external validation of the signal-processing front end.
  • VI. CONCLUSION: Attention fusion was strongest generally but failed under severe single-modality ECG degradation, while adaptive gates detected complete loss but scarcely tracked graded signal quality.These results distinguish modality availability from quality when both signals remain present.
Loading 2608.26000v1…