"Wrist-worn PPG segments"

"Photoplethysmography (PPG) from wearables offers a non-invasive, scalable alternative to electrocardiography (ECG) for continuous cardiac rhythm monitoring, with promising applications in atrial fibrillation detection. However, existing methods often rely on limited irregularity features, focus on single arrhythmia types, and lack validation across diverse populations, restricting clinical utility. We propose an interpretable machine learning framework using comprehensive feature engineering to detect multiple arrhythmias from PPG signals. Our study integrated four heterogeneous datasets: self-collected wrist-worn data, MIMIC-III, PhysioNet\/CinC 2015, and a GitHub PPG arrhythmia repository, yielding 72,442 ten-second segments from wrist, fingertip, and ICU devices across varied patient cohorts. A unified binary classification (normal versus arrhythmia) encompassed premature contractions, atrial fibrillation, ventricular tachycardia, and supraventricular tachycardia to support clinical triage. We extracted 85 physiologically meaningful features across six domains (time, HRV, frequency, morphology, physiology, and signal properties), then selected 50 discriminative features using SelectKBest (ANOVA F-statistic). Evaluating eight classifiers on stratified 70\/15\/15 splits, XGBoost achieved optimal test performance (n=10,867): 95.51% F1-score, 94.93% accuracy, 96.78% precision, 94.28% recall, and 99.22% ROC-AUC. Feature importance analysis revealed that robust arrhythmia detection requires multi-domain integration, particularly HRV entropy, pulse morphology dynamics, frequency descriptors, and signal quality metrics. Cross-dataset validation confirmed consistent performance across devices and populations, enhancing generalizability and mitigating dataset-specific bias for real-world deployment."