Status-Aware Self-Supervised Forecasting for Irregular Clinical Time Series
Status-Aware Self-Supervised Forecasting for Irregular Clinical Time Series
Kwanhyung Lee, Joohyung Lee, Jong-Heon Kim, Sangchul Hahn, Eunho Yang
Electronic health record (EHR) time series are irregular event streams, and many clinical tasks require forecasting a patient’s future state from their history. Because labeled outcomes in EHR are often limited and class-imbalanced, self-supervised learning (SSL) on large unlabeled cohorts is attractive. However, most existing EHR SSL methods first convert irregular event streams into discretized time grids, thereby losing the native event-set structure of observations. We propose a downstream task-aligned pretraining framework that models EHR trajectories as event sets and learns representations by forecasting future clinical states in latent space from the full past context. Our framework combines a momentum teacher for stable targets, a query-based Transformer decoder to predict variable future event sets, and an auxiliary masked event objective to improve local robustness. Across multiple ICU prediction tasks, our approach consistently improves downstream performance.
