<![CDATA[This study proposes a multimodal self-supervised framework for early detection of cardiac arrhythmias based on wearables combining 1-lead ECG, PPG, and IMU. The core method includes contrastive pretraining + masked reconstruction on synchronized windows and adaptive fusion weighted by Signal Quality Index (SQI) and aleatoric uncertainty, complemented by domain adaptation for invariant representation across devices and populations. The unlabeled corpus for pretraining contains 2,400 hours of free-living data from 820 participants (three different devices), while fine-tuning and clinical testing used 1,100 hours of labeled data (n=210; paroxysmal AF, PVC/PAC, SVT, episodic brady/tachycardia). In subject-wise testing, the model achieved Se 92.8%, Sp 97.1%, F1 90.3%, AUROC 0.972 for AF; F1 83.6% for PVC/PAC; and Se 88.9% for SVT. At episode-level evaluation (≥30 s), AF sensitivity was 94.6% with false alarms per hour (FPh) of 0.28 and a median time-to-detection of 22 s. Robustness increased at high activity (ECE 0.032, NLL −27%), leave-device-out generalization remained strong (AUROC 0.957), and the on-device implementation met resource limits (~68 ms/window on an edge-class MCU, <2.3 MB memory). These results demonstrate that signal quality/uncertainty-aware multimodal SSL can suppress false alarms without sacrificing sensitivity, enabling reliable and label-efficient home monitoring for wearable-based arrhythmia screening.]]>
Copyrights © 2025
