Cluster-Guided Self-Supervised Learning with Triplet Refinement for Longitudinal Heart Rate Time Series

The growing volume of longitudinal health data from wearable devices enables continuous monitoring and personalized insights. However, most of this data lacks reliable annotations, making manual labeling costly. In this context, clustering offers a practical solution by uncovering latent physiological patterns - such as sleep, exercise, and awake states - without requiring ground truth. To support clustering-based behavior discovery, we propose a self-supervised learning framework for physiological time series. Our method combines contrastive learning with a K-Means-guided clustering loss to encourage semantically coherent embeddings, and applies a lightweight post-hoc triplet loss to refine cluster boundaries. We evaluate our approach on a multi-subject dataset with a duration of a full day (24 hours) and a multi-year wearable heart rate single-subject dataset, demonstrating improvements over baseline methods in both clustering and classification metrics. Notably, the model maintains strong performance under high rates of input dropout, highlighting its robustness to real-world sensor noise and data sparsity. These results support the framework's potential to enable coarse-grained behavioral monitoring.