Explainable machine learning for 2-year functional outcomes in older adults with aneurysmal subarachnoid hemorrhage in North China: external validation and health-system context.

BACKGROUND

Older adults with aneurysmal subarachnoid hemorrhage (aSAH) have heterogeneous long-term outcomes, and differences in treatment pathways and hospital capacity may complicate early risk stratification.

METHODS

This multicenter cohort study included adults aged ≥60 years with aSAH in North China. Of 1,132 admissions at four hospitals from January 2017 to September 2023, 914 patients with 2-year follow-up were divided 7:3 into training (n = 640) and internal-validation (n = 274) cohorts. An independent cohort screened 602 admissions from January 2017 to January 2021 and included 499 patients. Poor outcome was defined as modified Rankin Scale 3-6 at 2 years. Least absolute shrinkage and selection operator regression with 10-fold cross-validation selected predictors, and 10 machine-learning algorithms were compared. Discrimination, calibration, decision-curve analysis, and SHapley Additive exPlanations (SHAP) were evaluated.

RESULTS

Among 1,413 patients, 583 (41.3%) had a poor 2-year outcome. Gradient boosting machine (GBM) AUCs were 0.863 (95% CI, 0.835-0.891), 0.834 (0.785-0.883), and 0.791 (0.749-0.833) in the training, internal-validation, and external-validation cohorts, respectively. External-validation accuracy was 0.752, sensitivity 0.661, specificity 0.808, precision 0.683, and F1 score 0.672. GBM had the highest external AUC point estimate but did not lead every classification metric. SHAP ranked Hunt-Hess grade, age, World Federation of Neurosurgical Societies grade, and treatment as the most influential features. Unadjusted survival differed across treatment and hospital-capacity strata (both log-rank p < 0.0001).

CONCLUSION

An explainable GBM retained moderate discrimination during external validation and may support admission-stage risk stratification in older adults with aSAH. The health-system findings are associative, and prospective recalibration, fairness assessment, and impact evaluation are required before clinical or public-health deployment.