Construction and Validation of a 90-Day Mortality Risk Prediction Model Based on Interpretable Machine Learning for Acute Ischemic Stroke Patients Undergoing Mechanical Thrombectomy

Background/Objectives: The accurate prediction of postprocedural mortality is critical for clinical decision-making; however, research on mortality risk models for patients undergoing mechanical thrombectomy remains limited. This study aimed to develop and validate machine learning models for predicting 90-day post-mechanical thrombectomy mortality. Methods: A retrospective-prospective cohort study involving 699 retrospective patients (January 2019–December 2022) and 274 prospective patients (January 2023–June 2024) from a single institution in Sichuan was conducted. The primary outcome was all-cause mortality within 90 days, ascertained via telephone follow-up. Predictors were identified using univariate analysis and LASSO regression. Eight predictive models were developed and evaluated using existing machine learning methods via 10-fold cross-validation. Model performance was assessed through discrimination, calibration, decision curve analysis, and interpretability via Shapley additive explanations. Results: The final dataset included 593 patients in the modeling set and 247 in the validation set. The 90-day mortality rates were 25.6% and 32.0%, respectively. Key predictors included age, hyperlipidemia, atrial fibrillation, pre-stroke statin use, antiplatelet/anticoagulant therapy within 48 h of onset, dysphagia, D-dimer levels, and activities of daily living scores. Logistic regression demonstrated superior performance in the modeling cohort (AUC = 0.87), whereas the multilayer perceptron model exhibited the greatest efficacy in the validation cohort (AUC = 0.77). Conclusions: Machine learning algorithms can accurately predict 90-day mortality among patients undergoing mechanical thrombectomy. The multilayer perceptron model demonstrated robust validation performance and offers a potential tool for personalized risk assessment and optimization of clinical decision-making.