Development and validation of an interpretable machine learning model for predicting hospital mortality in...

To develop and validate an interpretable machine learning model for predicting hospital mortality in patients with ovarian cancer admitted to intensive care units (ICUs).

This retrospective multicenter study analyzed 433 patients with ovarian cancer from MIMIC-IV (n = 286) and eICU (n = 147) databases who met inclusion criteria of first ICU admission, aged >18 years, and ICU stay ≥24 hours. The Boruta algorithm identified important predictors from clinical and laboratory data collected within 24 hours of admission. Seven machine learning algorithms were evaluated using 10-fold cross-validation. MIMIC-IV served as the training data set (70% training, 30% internal validation) with eICU providing external validation. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) analysis to identify feature contributions and clinical thresholds.

The final cohort had 9.5% hospital mortality (41/433). The Support Vector Machine model achieved superior performance with area under the curve of 0.857 (95% CI 0.737 to 0.976) in internal validation and 0.750 (95% CI 0.642 to 0.858) in external validation, outperforming Sequential Organ Failure Assessment (0.578), Simplified Acute Physiology Score II (0.674), and Oxford Acute Severity of Illness Score (0.671) scores. Red cell distribution width (RDW) emerged as the most important predictor (mean importance 0.032), followed by bicarbonate (0.028) and chloride (0.020). SHAP analysis revealed critical thresholds: RDW >20% (particularly >25%), bicarbonate <20 mmol/L, and abnormal chloride levels significantly increased mortality risk. High-risk patients demonstrated RDW 30.8% (SHAP +0.675) and bicarbonate 12 mmol/L (SHAP +0.124).

The interpretable machine learning model accurately predicts hospital mortality in ICU-admitted patients with ovarian cancer using readily available laboratory parameters, significantly outperforming traditional ICU scores, and providing actionable clinical thresholds for risk stratification.