Quality-Aware Multimodal Classification of Synchronized Ecg and Ppg: A Critical Comparison of Algorithms and a Reproducible Experimental Framework

Synchronized electrocardiography (ECG) and photoplethysmography (PPG) provide complementary infor-mation about cardiac electrical activity and peripheral blood-volume pulsation. This study critically compares classical machine-learning classifiers, tree ensembles, convolutional neural networks, recurrent architectures, and attention-based models for multimodal biosignal classification. A quality-aware ECG–PPG fusion method (QAEF) is proposed, combining modality-specific feature representations, signal-quality-weighted fusion, cross-modal timing descriptors, and uncertainty-based abstention. The methodological analysis uses the BIDMC PPG and Respiration Dataset, which contains 53 synchronized recordings of approximately eight minutes each, sampled at 125 Hz. However, the dataset does not provide ready-made expert arrhythmia labels. Therefore, a reproducible exploratory experiment is defined around monitor-derived disagreement between heart rate (HR) and pulse rate (PULSE), rather than clinical diagnosis. The protocol specifies record-wise data splitting, unimodal and multimodal baselines, evaluation metrics, and sanity checks. Published findings are explicitly separated from results that would require execution of the experiment. Since the raw data could not be downloaded and evaluated in the current environment, no new accuracy or F1-score is claimed. QAEF re-mains a research hypothesis requiring patient-wise validation, controlled signal-degradation tests, and uncer-tainty calibration.