Explainable machine learning model for predicting the outcome of acute ischemic stroke after intravenous thrombolysis

Abstract

Introduction: Acute ischemic stroke (AIS) patients often experience poor functional outcomes post-intravenous thrombolysis (IVT). Novel computational methods leveraging machine learning (ML) architectures increasingly support medical decision-making. We aimed to develop and validate a machine learning model to predict 3-month unfavorable functional outcome after IVT in AIS patients. Methods: This retrospective study developed ML prognostic models for 3-month functional outcome (modified Rankin scale scores of 3–6) in IVT-treated AIS patients. A derivation cohort (n = 938) was split 7:3 for training/testing, with an independent external validation cohort (n = 324). The least absolute shrinkage and selection operator (LASSO) regression selected predictors from clinical/neuroimaging/laboratory variables. Eight ML algorithms (including Logistic Regression, Random Forest, Extreme Gradient Boosting, Multilayer Perceptron, Support Vector Machine, Light Gradient Boosting Machine, Decision Tree, and K-Nearest Neighbors) were trained using 10-fold cross-validation and evaluated on test/external sets via the area under the curve (AUC), accuracy, precision, recall and F1-score. Additionally, the SHapley Additive exPlanations (SHAP) interpreted the optimal model. Results: 938 patients constituted the derivation cohort (training: n = 656, test: n = 282) and 324 patients the external validation cohort. Unfavorable 3-month outcomes (mRS 3–6) occurred in 25.7% and 22.8%, respectively. LASSO regression selected five predictors: the neutrophil-to-lymphocyte ratio (NLR), admission National Institutes of Health Stroke Scale (NIHSS) score, the Alberta Stroke Program Early CT Score (ASPECTS), atrial fibrillation, and blood glucose. While tree-based methods like XGBoost and LightGBM showed elevated training performance (e.g., XGBoost training AUC = 0.878) but significant drops in validation (AUC = 0.791), LR demonstrated optimal performance: robust training AUC (0.792), minimal validation degradation (AUC = 0.787). LR model was subsequently employed as classification method demonstrating optimal performance with (AUC = 0.777) in the test dataset. External validation confirmed LR’s stability (AUC = 0.797). SHAP analysis ranked NLR as the strongest predictor (followed by NIHSS/ASPECTS), with higher values increasing risk. Learning curves indicated no overfitting. A nomogram enabled individualized risk quantification. Conclusion: A parsimonious 5-variable LR model robustly predicts 3-month post-IVT outcomes, combining clinical utility, interpretability, and generalizability. NLR-driven inflammation is critical to prognosis. This tool facilitates early high-risk patient identification for personalized intervention.