Calibration-Aware Multimodal Forecasting of ICU Deterioration Under Irregular Sampling and Documentation Sparsity

Abstract

Intensive care units produce high-velocity streams of structured physiologic measurements and lower-frequency clinical documentation, both of which can contribute to early warning systems for impending deterioration. While modern deep sequence models often improve ranking performance for near-term adverse events, they frequently output probabilities that are poorly calibrated, especially when trained under extreme class imbalance and heterogeneous missingness. In operational settings, calibration is not a cosmetic property: it governs alarm thresholds, resource allocation, and downstream decision policies that interpret scores as risk. This paper develops a calibration-aware, multimodal forecasting framework for ICU deterioration prediction that explicitly couples representation learning with probability quality constraints. We formalize hourly prediction as conditional risk estimation over partially observed multivariate histories augmented by asynchronous text context, and we analyze how common optimization choices for imbalanced learning can distort probabilistic semantics even when discrimination remains strong. We then propose a structured methodology that separates discrimination learning from calibration recovery through principled post-hoc maps, while maintaining strict temporal availability to avoid leakage from documentation latency. The approach combines mask- and staleness-aware temporal encoders, modality-stable fusion operators, and calibration layers designed to preserve ordering while repairing probability scale. We further outline validation protocols that measure calibration under prevalence shift, quantify reliability under missingness stratification, and translate probability error into operational utility under alerting constraints. The result is a technically grounded blueprint for building multimodal ICU risk models whose outputs can be meaningfully used as probabilities rather than merely as scores.