Development and validation of interpretable machine learning models for predicting long-term functional outcomes in elderly patients with aneurysmal subarachnoid hemorrhage.

BACKGROUND Accurate prognostication in elderly patients with aneurysmal subarachnoid hemorrhage (aSAH) remains challenging due to high morbidity and mortality. This study aimed to develop and externally validate interpretable machine learning (ML) models for predicting 12-month functional outcomes.

METHODS Data from 426 consecutive elderly aSAH patients at the primary center were randomly split into a training cohort (n = 298, 70%) and an internal test cohort (n = 128, 30%). An independent external validation cohort (n = 41) was obtained from a collaborating tertiary medical center. Features with multicollinearity (|Spearman ρ| > 0.8) were excluded. Predictive variables were identified using univariate/multivariate logistic regression and the Boruta algorithm. Eight ML models were trained using 5-fold cross-validation. Model performance was assessed on both internal and external validation cohorts using receiver operating characteristic (ROC) and precision-recall (PR) curves, along with calibration plots. Model interpretability was evaluated using SHapley Additive exPlanations (SHAP).

RESULTS The Multilayer Perceptron (MLP) model demonstrated superior performance, achieving ROC-AUCs of 0.913 (internal testing) and 0.912 (external validation), with favorable calibration in both cohorts. SHAP analysis identified the Hunt-Hess scale, age, total bleeding volume, delayed cerebral ischemia (DCI), modified Fisher scale, rebleeding, hydrocephalus, aneurysm multiplicity, and aneurysm length as key predictors. SHAP dependency plots facilitated individualized risk interpretation.

CONCLUSION This study successfully developed and externally validated interpretable ML models that reliably predict long-term functional outcomes in elderly aSAH patients. These tools demonstrate robust generalizability across clinical settings and hold potential to support personalized clinical decision-making and optimize resource allocation in neurocritical care.