An interpretable machine learning model for predicting prognosis in acute ischemic stroke with large vessel occlusion

Abstract

Objective To develop and validate an interpretable nomogram that integrates an imaging score, clinical characteristics, and machine-learning models to predict 90-day functional outcomes in patients with acute ischemic stroke (AIS) and large-vessel occlusion (LVO). Methods We analyzed 240 AIS patients with anterior circulation LVO who underwent one-stop multimodal CT between October 2019 and December 2024. Fifty-two variables, including pretreatment clinical features, conventional and advanced imaging, and angiographic characteristics, were assessed. The 90-day modified Rankin Scale (mRS-90) was used as the prognostic endpoint; good and poor outcomes were defined as mRS-90 ≤ 2 and > 2, respectively. Patients were randomly assigned to training (80%, n = 192) and testing (20%, n = 48) cohorts. Least absolute shrinkage and selection operator (LASSO) regression was used to derive the imaging score, which was then combined with key clinical predictors to construct a nomogram. Clinical, imaging, and hybrid models were developed separately and evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). Repeated stratified 5-fold cross-validation repeated 20 times was performed as a stability analysis of the fixed final model structures. SHapley Additive Explanations (SHAP) was applied to identify influential features and visualize feature importance and interactions. Results Among 240 patients, 162 (67.5%) had unfavorable 90-day functional outcomes. Patients with unfavorable outcomes were generally older, had higher admission NIHSS scores, and showed less favorable collateral and perfusion profiles. Age, admission NIHSS score, and onset-to-CT time were retained as clinical predictors, while eight imaging features were integrated into the imaging score. Adding the imaging score did not significantly improve discrimination in the testing cohort (ΔAUC, 0.0078; 95% CI, −0.0071 to 0.0227; P = 0.305). In repeated stratified 5-fold cross-validation performed 20 times, the mean AUCs were 0.942, 0.747, and 0.939 for the clinical, imaging, and hybrid models, respectively. SHAP analysis characterized the contributions of age, NIHSS, onset-to-CT time, and the composite imaging score within the four-input hybrid model. Conclusions An interpretable model combining routine clinical predictors with a multimodal CT-derived imaging score demonstrates promising predictive potential for 90-day functional outcome in patients with AIS-LVO.