Missingness-Aware Temporal Modeling of Electronic Health Records for Early Clinical Deterioration Prediction

Abstract

Electronic health records contain irregularly sampled measurements and substantial missing data, creating difficulties for models that predict patient deterioration from longitudinal observations. This paper develops a missingness-aware temporal model that explicitly incorporates both observed clinical values and the timing of unavailable measurements. For each variable, the model jointly represents its latest observed value, elapsed time since measurement, and recent observation frequency. These representations are processed through a temporal prediction network to estimate deterioration risk within a predefined clinical window. Unlike conventional preprocessing pipelines that treat missing values primarily as a data quality problem, the proposed method allows measurement patterns themselves to contribute to prediction when they reflect changes in clinical attention or patient condition. Evaluation on longitudinal hospital records shows improvements in AUROC, AUPRC, and early-warning sensitivity over standard imputation-based temporal models. Ablation experiments indicate that measurement intervals and observation frequency provide complementary information to physiological values, especially during the hours preceding deterioration.