Self-Supervised Learning Improves Early Sepsis Prediction from EHRs.

This research introduces a novel framework for predicting sepsis early using electronic health records (EHRs), addressing common data issues like irregular sampling and missing information. The approach leverages self-supervised learning techniques, specifically Joint Embedding Predictive Architecture (JEPA) and VICReg, alongside federated representation learning. By systematically comparing various modeling paradigms, the study identified that self-supervised pretraining, particularly with VICReg followed by semi-supervised fine-tuning, yielded robust and temporally persistent representations. The most effective model, combining JEPA with XGBoost and mean pooling, achieved a high AUPRC score for sepsis onset prediction, nearly matching a leading benchmark while utilizing 83% fewer biomarkers. Furthermore, the VICReg-based pipeline significantly improved upon raw-feature baselines and end-to-end supervised methods, demonstrating its ability to generate features that are both accurate at onset and stable over longer prediction horizons. This suggests a powerful method for extracting meaningful insights from complex medical data.