Development and validation of a machine learning model to predict functional outcomes in patients with recent small subcortical infarction

OBJECTIVE: A substantial proportion of patients (12 %-25 %) with recent small subcortical infarction (RSSI) suffer poor functional outcomes at 3 months. Despite the identification of prognostic factors, a significant gap exists in predictive modeling. This study aimed to develop and validate machine learning models to accurately predict 3-month functional status in this patient population. METHODS: This multicenter study prospectively enrolled 1576 patients diagnosed with RSSI. The primary cohort (n = 1126) was randomly split into a training set (70 %) and an internal validation set (30 %). An independent external cohort (n = 450) was used for further validation. The primary outcome was an unfavorable functional status at 3 months, defined as a modified Rankin Scale (mRS) score ≥3. The Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression model was employed for feature selection from demographic, clinical, laboratory, and imaging variables. 8 supervised machine learning models were developed and compared. Model performance was rigorously evaluated in the validation cohorts using the Area Under the Receiver Operating Characteristic Curve (AUC) for discrimination, calibration curves for consistency, and Decision Curve Analysis (DCA) for clinical utility. The optimal model was interpreted using SHapley Additive exPlanations (SHAP). RESULTS: LASSO regression identified 8 features with non-zero coefficients for predicting outcomes: NIHSS, proximal RSSI (pRSSI), glucose, stress hyperglycemia ratio (SHR), neutrophil-to-lymphocyte ratio (NLR), age, systolic blood pressure (SBP), and LDL-C. Among the eight developed models, the CatBoost model demonstrated the best performance. It achieved the highest AUC in the training set (0.961), the internal validation cohort (0.940), and the external validation cohort (0.875). The CatBoost model also showed excellent calibration and provided the greatest net benefit across a wide range of threshold probabilities in DCA for both validation cohorts. SHAP analysis identified the NIHSS score as the most significant predictor of unfavorable outcomes, followed by pRSSI, Glucose, SHR, and NLR. A publicly accessible web tool based on the model is available at: https:// predictrssi.streamlit.app. CONCLUSION: This study successfully developed and validated a robust CatBoost machine learning model that accurately predicts 3-month functional outcomes in patients with RSSI using eight readily accessible features. This model, which outperforms seven other machine learning algorithms, is available as a user-friendly web application to aid clinicians in early risk stratification and personalized patient management.