Semi-Supervised Contrastive Learning for Time Series Classification in Healthcare

Healthcare data comprises a diverse array of information, such as physiological indicators, activity behavior, sleep quality, and emotional state, among others. These features provide valuable insights into individuals' health status, behavioral patterns, and quality of life. Nevertheless, the lack of readily identifiable labeling information in this data poses a significant challenge for learning, particularly with label-limited time series data. Traditional annotation methods are prohibitively costly and time-consuming, making them impractical for large-scale applications. To tackle this challenge, this paper presents a novel approach called semi-supervised contrastive learning for time Series classification in healthcare (SSC-TC). This approach leverages existing positive-negative pairs and introduces new selection rules and loss functions to enhance the learning process. By generating pseudo-labels on the initially label-restricted dataset, the proposed model extracts more meaningful label information and learns richer representations. Experimental results demonstrate the superiority of the SSC-TC over other baseline methods, both on balanced and imbalanced datasets. The effectiveness of the approach is further validated through ablation experiments, where different components of the model are evaluated and their contributions to the overall performance are assessed. Overall, this research highlights the potential of semi-supervised contrastive learning in overcoming the limitations of label-restricted time series data in the field of healthcare. It offers a promising solution for leveraging the vast amount of wearable device data and extracting valuable insights for personalized caregiving without the need for extensive manual annotation.