Graph-Enhanced Cross-Domain Robust Contrastive Learning for Noisy and Disparate Medical Time Series Analysis

Medical time series data, such as electroencephalograms and electrocardiograms, are vital for diagnosis but face challenges from noise, cross-domain variability, and limited labeled data. Traditional and existing contrastive learning methods often struggle to yield robust and generalizable models. We propose Graph-Enhanced Cross-domain Robust Contrastive Learning (GCRoCL), a novel framework for learning noise-robust and domain-invariant representations from medical time series. GCRoCL integrates dynamic graph construction and graph-level data augmentation to capture complex relationships and enhance resilience to noise. A hybrid Graph-Temporal Feature Encoder extracts comprehensive spatiotemporal features. A Cross-domain Adversarial Contrastive Learner then uses contrastive loss and a Domain Adversarial Alignment module to enforce discriminative and domain-invariant feature learning. Extensive experiments demonstrate GCRoCL's superior performance, consistently outperforming state-of-the-art baselines. It also exhibits strong cross-domain generalization and remarkable robustness in low-labeled data regimes. Ablation studies confirm the vital contribution of its core modules. These results underscore GCRoCL's potential for reliable and generalizable diagnostic tools in diverse clinical settings.