A contrastive learning approach for time series with application to anomaly detection in healthcare

Time-series analysis, particularly in biomedical signals like Electrocardiograms (ECG), poses the critical challenge of distinguishing between global patterns shared across subjects and individual-specific features unique to each patient. Many existing methods fail to achieve this balance, limiting their applicability to tasks requiring both generalization and fine-grained differentiation. This thesis addresses this limitation by exploring the central research question: How can contrastive learning techniques help distinguish general patterns from individual fingerprints in time-series data? To answer this, a contrastive learning framework is proposed, leveraging advanced preprocessing and deep learning techniques. Seasonal-Trend decomposition using Loess (STL) is employed to preprocess ECG signals, isolating periodic components while preserving critical details. The proposed Siamese network architecture, enhanced with Residual Blocks and Squeeze-and-Excitation (SE) mechanisms, generates robust embed- dings that effectively balance global pattern extraction and individual fingerprint preservation. The research question is evaluated through contrastive loss, t-SNE visualizations of the embedding space, and L2 distance distributions for positive and negative pairs, which together assess the networks ability to capture intra-subject consistency and inter-subject variability. Anomaly detection is employed as a representative use case to further demonstrate the frameworks potential, with performance assessed using metrics such as accuracy, precision, recall, and AUC. While the focus of this work lies in addressing the research question, the methodologies and insights extend beyond the specific case of ECG anomaly detection. By bridging the gap between global pattern extraction and individual fingerprint preservation, this thesis provides a robust foundation for applying contrastive learning techniques to a wide range of time-series tasks.