Generalization, Cross-ICU Transfer, and Explainability of a Mortality and Time-to-Discharge Framework for the Intensive Care Unit.

Background

PADS combines two neural networks predicting ICU mortality and discharge within 48 h, placing critically ill patients into one of four clinically meaningful states. Developed on MIMIC-IV alone, it left open whether it generalizes to other ICUs, whether its models transfer across hospitals, and whether its predictions can be explained at the bedside.

Methods

We evaluated PADS on four ICU databases from different hospitals and countries (MIMIC-IV, AmsterdamUMCdb, eICU-CRD, and HiRID), using the same routinely collected variables. Mortality is scored on the final 48-h window (terminal-window, not early-warning, discrimination). For each external database, we compared the MIMIC model used as-is, retrained from scratch, and retrained from the MIMIC weights, and added an explainability layer.

Results

For mortality, reusing and retraining the MIMIC model gave the highest discrimination on every database (AUROC 0.955-0.986; terminal-window (near-outcome) discrimination) and stabilized training; used as-is, it ranged from chance (Amsterdam) to good (eICU, HiRID). For discharge, training fresh on local data matched or beat reusing MIMIC on every external database, consistent with discharge timing depending on local organization rather than physiology. The explainability layer produced clinically coherent, cross-checked explanations.

Conclusions

Transportability was task-dependent: mortality transferred between hospitals, discharge did not. PADS demonstrated promising external transportability across heterogeneous ICU databases, particularly after local adaptation. Reusing and adapting the MIMIC-IV mortality model across hospitals improves accuracy. This approach also stabilizes training, providing a basis for potential federated deployment, whereas discharge is better trained locally. The mortality results reported here are terminal-window discrimination and do not support use of the framework as an early-warning model. A transparent explainability layer provides an interpretable representation of model predictions, addressing a key barrier to clinical adoption.