MADE: leveraging informative missingness and temporal evaluation for electronic health record-based deterioration prediction in general wards

Abstract Objective Electronic health record (EHR) time series are irregularly sampled and frequently incomplete, with missingness potentially carrying clinically informative signals. We developed Masked-Attention-based timestamp-wise Data Embedding (MADE), a missingness-aware model for continuous prediction of deterioration in general wards, and evaluated it across heterogeneous hospitals. Materials and Methods Observed variables were tokenized into feature-specific embeddings and processed by a masked-attention Transformer, followed by bi-LSTM temporal aggregation for risk prediction. MADE was evaluated for major adverse event (MAE) and sepsis in 1 internal (SVH) and 2 external cohorts (AMC, HUMC) against conventional and deep learning baselines. Ablation studies compared no imputation with last observation carried forward (LOCF) and multiple imputations by chained equations (MICE). Discriminative performance was assessed using the area under the receiver operating characteristic curve (AUROC) and precision–recall curve (AUPRC). Clinical utility and reliability were assessed using decision curve analysis, Brier score, and expected calibration error (ECE). Results MADE achieved strong discrimination for MAE (AUROC 0.966/0.902/0.825 for SVH/AMC/HUMC) and sepsis (AUROC 0.934/0.874/0.776). The no-imputation configuration generally outperformed LOCF and MICE, particularly for MAE and for sepsis AUPRC, supporting preservation of informative missingness. Decision curves showed favorable net benefit for MAE and competitive benefit for sepsis. Calibration indicated modest underestimation at higher risks, suggesting site-specific recalibration. Discussion MADE mitigates limitations of fixed grid and imputation-dependent EHR modeling by jointly leveraging observed values and missingness patterns under heterogeneous sampling. Conclusion MADE supports scalable observation-aligned risk monitoring, particularly for MAE, while sepsis prediction may require site-specific recalibration or feature-set adaptation under cross-institution domain shifts.