Modeling Temporal Information Loss in Electronic Health Records: A Dual-Clock Approach to Multimodal Clinical Risk Prediction

Abstract

Digital risk modeling in hospitals is usually trained on variables as they appear in the record, even though the underlying signals often exist earlier in monitor feeds, device logs, text fragments, and other non-tabular streams. Once those sources are collapsed into charted fields, the learning problem becomes easier to standardize but poorer in time resolution. This study examines that loss from a data-science perspective by treating delayed chart visibility as a form of timestamp corruption. An empirical benchmark is constructed from a replay corpus of 58,432 encounter trajectories, 214.7 million multimodal observations, and nine institutional delay regimes. Each observation carries both an acquisition clock and a release clock, allowing direct comparison between models trained on raw availability and models restricted to record-time exposure. The proposed method, a dual-clock masked sequence architecture with release-aware contrastive regularization, is evaluated against lagged tabular baselines, real-time tabular transformers, sequence mixers, and multimodal encoders without clock constraints. Across leave-one-site-out evaluation, the proposed system achieves the strongest event-level precision-recall performance, the lowest calibration error, and the largest advance interval before downstream escalation. The empirical patterns are consistent across moderate prevalence shift, note dropout, image sparsification, and synthetic changes in documentation latency. When raw streams are forced to obey record-time release, most of the improvement disappears, indicating that the main advantage does not come from feature volume alone. It comes from preserving when evidence actually becomes observable. The study therefore reframes earlier-than-EHR modeling as a benchmark design problem centered on time semantics, representation robustness, and transport across institutions.