Development and external validation of an interpretable machine learning model for early prediction of in-hospital mortality in neurocritical care.

Background: Early risk assessment is a routine part of neurocritical care, yet mortality prediction remains difficult because neurological injury often coexists with respiratory, circulatory, renal, and inflammatory dysfunction. This study aimed to develop and externally validate an interpretable machine learning (ML) model for predicting in-hospital mortality in neurocritical-care patients using variables available within the first 24 h after intensive care unit (ICU) admission. Methods: We conducted a retrospective cohort study using two public critical-care databases. The Medical Information Mart for Intensive Care IV (MIMIC-IV) database was used for model development and internal testing, and the eICU Collaborative Research Database (eICU-CRD) was used for independent external validation. The report follows the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement. Adult ICU patients with neurocritical-care diagnoses were identified by predefined International Classification of Diseases codes. Candidate predictors were restricted to information recorded during the first 24 h after ICU admission and included demographics, comorbidities, vital signs, laboratory results, neurological status, organ-support therapies, and severity scores. Least absolute shrinkage and selection operator (LASSO) regression was used for feature selection. Logistic regression, random forest, XGBoost, LightGBM, CatBoost, and multilayer perceptron models were trained. Discrimination was assessed by area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC). Calibration, decision curve analysis (DCA), and SHapley Additive exPlanations (SHAP) were used to evaluate probability reliability, potential clinical utility, and interpretability. The operating threshold used for classification metrics was fixed in the MIMIC-IV development data before internal and external evaluation. Results: The MIMIC-IV cohort included Conclusions: A model based on routinely collected first-day ICU data provided externally validated and interpretable risk stratification for in-hospital mortality in neurocritical care. Its low positive predictive value at the reported operating point precludes use as a stand-alone basis for prognostic pessimism, treatment limitation, or withdrawal of life-sustaining treatment. Prospective validation in diverse countries and health systems is required before clinical implementation.