Time-Series Bert for Sepsis Detection: Uncovering Patient Trajectories Through Vital Sign Embeddings

This study adapts BERT for vital sign time-series analysis in sepsis detection. Using MIMIC-III data, our model's embeddings reveal patient clusters that partition septic from non-septic cases while capturing physiological complexity through diagnosis count distributions. The BERT-based classifier achieves robust performance in both Precision-Recall Area Under Curve (PR AUC), measuring precision maintenance across recall thresholds, and Receiver Operating Characteristic Area Under Curve (ROC AUC), quantifying septic/non-septic case discrimination. Unsupervised learning reveals patient subgroups with distinct physiological profiles, highlighting transformer architectures' ability to extract meaningful patterns from medical time-series for enhanced sepsis monitoring.