Toward transparent clinical decision support in the ICU: A multi-window, model-agnostic explainability approach for...

This study has two main objectives: (1) to develop a multi-model framework for predicting Intensive Care Unit (ICU) mortality within the first 72 h of admission; and (2) to introduce a novel model-agnostic explainability approach classification that enables variable-level interpretation of predicted probabilities.

Retrospective study using a multi-model machine learning approach, analyzing data across multiple time windows and incorporating demographic, clinical, and biochemical variables.

ICU mortality.

Patients included in the eICU database over 16 years old who have been admitted to ICUs in 2014 and 2015 with available data within the first 72 h after ICU admission. A total of 106,449 patients were included in the analyses.

No clinical interventions were applied; this was a retrospective analysis for predictive model development and evaluation.

Demographic, clinical, and biochemical variables collected across multiple time windows.

A total of 106,449 were included (mean age 62.6 years, 46% women), with an overall 72-h mortality of 4.8%. Random Forest models achieved one of the best results in terms of predictive performance metrics, with F1-scores of 0.93, 95% CI 0.93 to 0.94; 0.92, 95% CI 0.92 to 0.93 and 0.83, 95% CI 0.83 to 0.85 across the three temporal data windows. Due to these metrics, the ability to predict deaths, and the biological plausibility of the predictions, Random Forest models were selected from all those studied.

The proposed multi-model approach significantly improves 72-h ICU mortality prediction. Moreover, we outline a model-agnostic strategy for variable-level interpretation of predicted probabilities, which may facilitate transparency and support future applications in clinical decision support.