An attention-based model that combines ECG and PPG had the lowest overall heart-rate error in a controlled test of degraded signals, but its apparent advantage was not confirmed after correction for multiple comparisons. Across six degradation regimes, its mean absolute error, or MAE, was 1.66 ± 0.43 beats per minute across five training seeds.
A controlled comparison
The analysis covered three pre-registered architecture comparisons across seven metrics, making 21 tests. Models were trained independently under five random seeds, numbered 0 through 4, on a fixed shared train, validation and test split. The researchers used exact paired Wilcoxon signed-rank and paired t-tests, effect sizes and t-based 95% confidence intervals, followed by Holm–Bonferroni correction.
Researchers compared eight ECG–PPG fusion architectures with an identical encoder backbone and training protocol. The comparison included unimodal estimates, fixed fusion, attention fusion and adaptive gating, including a version conditioned on signal-quality information. The central question was whether these approaches would remain accurate as one or both inputs became less reliable or disappeared entirely.
The systems were tested on synthetic eight-second windows sampled at 125 Hz, with target heart rates drawn from 50 to 110 bpm. Five corruption levels were collapsed into six analysis regimes. The synthetic data included 792 training pairs, 234 validation pairs and 324 test pairs, chosen for single-CPU tractability and at least about 20 test windows per regime cell rather than as an optimized sample size.
Across the full set of regimes, the implicit gate had an overall MAE of 2.06 ± 0.51 bpm and the SQI-conditioned gate had an MAE of 1.99 ± 0.61 bpm. Fixed-weight and unimodal methods were descriptively worse, with errors ranging from 2.18 to 2.87 bpm. These were descriptive rankings, not corrected pairwise findings.
The edge changed with the failure mode
The overall ranking did not hold across every type of damage. Attention fusion was strongest at every graded level of PPG degradation, but under severe ECG degradation, PPG-only estimation had an MAE of 1.45 bpm, compared with 2.15 bpm for attention fusion.
Complete sensor loss produced a sharper split. When PPG was missing, ECG-only estimation had an MAE of 1.48 bpm, while the SQI-conditioned gate recorded 1.56 ± 0.59 bpm, within 0.08 bpm of the unimodal ceiling. When ECG was missing, PPG-only estimation had an MAE of 1.47 bpm, but the SQI-conditioned gate rose to 4.36 ± 1.59 bpm. The unimodal methods were much less accurate when their own signal was absent: ECG-only reached 14.34 bpm with missing ECG, and PPG-only reached 15.02 bpm with missing PPG.
The real-data check was narrower
Separate real-data checks were used to validate the signal-processing front end: five fetal-ECG records, 53 ICU patients, 848 30-second windows and 7,868 matched beats for PTT analysis. On direct fetal leads, R-peak F1 ranged from 0.89 to 0.98, with a mean of 0.95. Naive abdominal-lead detection was 0.19 to 0.34 and correlated with the signal-quality pass rate at r = 0.927. In the ICU data, ECG-derived rate MAE was 1.61 bpm, compared with 2.78 bpm for PPG-derived pulse rate; correlations with the bedside reference were 0.911 and 0.879, respectively.
The gates saw absence more clearly than quality
Both adaptive gates allocated more than 98% of their weight to the modality that survived complete dropout. But when both signals were still present and only graded-degraded, the correlation between ECG gate weight and the signal-quality difference was 0.24 for the implicit gate and 0.10 for the SQI-conditioned gate. When all conditions were pooled, those correlations were 0.68 and 0.70. The pattern is consistent with strong detection of all-or-nothing absence, but limited tracking of quality differences while both channels remained available.
An early result with a narrow reach
The fusion comparison itself was based on physiologically grounded synthetic window pairs, while real recordings were used to validate the signal-processing front end. The findings therefore describe comparative behavior under the programmed degradation regimes; they do not establish clinical or real-world deployment performance. With only five seeds, the closest exact Wilcoxon p-value was 0.0625, and no pre-registered comparison survived Holm correction.
The supplied document is an arXiv v1 preprint dated 26 Aug 2026 and says the manuscript was submitted for review. It reports that signal-processing and fusion-architecture code is available in a project repository, while synthetic-data generation code, evaluation scripts and full seed-level results tables are provided as supplementary material. No funding or conflict-of-interest disclosure is reported in the supplied text.
Paper data and sources
Original title: CardioFusion-AI: Robust ECG--PPG Fusion for Multimodal Physiological Monitoring Under Signal Degradation
Authors: Navaneetha Krishnan Kamalakannan, Janakiraman Kamalakannan
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text