Abstract
OBJECTIVE: To develop and validate explainable machine learning models for predicting 3-month functional outcomes in acute ischemic stroke (AIS) patients using SHapley Additive exPlanations (SHAP) framework. METHODS: This retrospective cohort study included 538 AIS patients admitted within 72 h of symptom onset. Patients were randomly divided into training (70%) and validation (30%) sets. Clinical, laboratory, and imaging data were collected. Least Absolute Shrinkage and Selection Operator regression was used for feature selection. Five machine learning models were developed: support vector machine, k-nearest neighbors, random forest, gradient boosting machine (GBM), and convolutional neural network. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. SHAP analysis was applied to the best-performing model to enhance interpretability. RESULTS: Among 538 patients (mean age 68.5 ± 12.7 years, 58.0% male), 34.2% had poor 3-month outcomes (mRS 3-6). The GBM achieved the best predictive performance with AUC of 0.91, accuracy of 0.81, sensitivity of 0.95, and specificity of 0.61 in validation set, significantly outperforming logistic regression (AUC = 0.78). The model demonstrated excellent calibration and superior net benefit in decision curve analysis across threshold probabilities of 0.1-0.7. SHAP analysis identified admission NIHSS score (30.8%), age (14.9%), and ASPECTS ≥7 (13.7%) as the most influential predictors, with neutrophil-to-lymphocyte ratio (10.1%) and platelet distribution width (9.7%) also contributing significantly to outcome prediction. CONCLUSION: Explainable machine learning models can accurately predict 3-month functional outcomes in AIS patients. The SHAP framework enhances model transparency, addressing interpretability barriers for clinical implementation while maintaining superior predictive performance.