Domain-Adaptive Arrhythmia Classification on Wearable Device Data
Abstract
Long-term electrocardiogram (ECG) monitoring is crucial for detecting intermittent arrhythmias, and wearable devices enable such monitoring in ambulatory settings. However, robust classification of wearable-device ECG recordings remains challenging because labeled wearable-device ECG data are far less abundant than hospital ECG data and differ substantially in their statistical distributions. We propose CaCTuS (Class-aware Contrastive Learning with Temporal mixup under Supervision), a supervised multi-domain framework designed to address the substantial disparity in data availability between hospital and wearable-device domains. CaCTuS generates mixed samples through intra-class temporal mixup while interpolating dual domain-specific CLS tokens with the corresponding mixup coefficient. This approach retains domain-specific information and provides semantically consistent supervision for cross-entropy and supervised contrastive learning, facilitating robust, class-discriminative representation learning across domains. On our wearable-device (MEMO Patch) test set, CaCTuS achieves a mean F1 score of <inline-formula> <tex-math notation="LaTeX">$79.21\pm 0.51\%$ </tex-math></inline-formula>, outperforming the next-best baseline by 4.98 percentage points. On the hospital (SNU-BH) test set, CaCTuS attains <inline-formula> <tex-math notation="LaTeX">$91.53\pm 0.62\%$ </tex-math></inline-formula>, matching the strongest hospital-only baseline (<inline-formula> <tex-math notation="LaTeX">$91.86\pm 0.31\%$ </tex-math></inline-formula>) within 0.33 percentage points. To assess generalizability beyond our internally collected hospital data, we further evaluate CaCTuS on publicly available MIT-BIH data under our label taxonomy. CaCTuS achieves F1 scores of <inline-formula> <tex-math notation="LaTeX">$79.94\pm 0.93\%$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$93.13\pm 1.49\%$ </tex-math></inline-formula> on the wearable-device (MEMO Patch) and hospital (MIT-BIH) domains, respectively, exceeding the next-best baselines by 22.92 and 2.19 percentage points.