An Explainable Machine Learning Model Integrating Transcranial Doppler and Clinical Data to Predict Outcomes After...

Endovascular thrombectomy (EVT) has improved outcomes in patients with acute ischemic stroke due to large vessel occlusion; however, a substantial proportion of patients still experience poor functional outcomes despite successful recanalization. Growing evidence suggests that postprocedural cerebral hemodynamics are closely associated with recovery after EVT. This study aimed to develop an interpretable machine learning model integrating transcranial Doppler (TCD)-derived hemodynamic parameters and clinical features to predict 3-month functional outcomes after EVT.

This retrospective cohort study included 176 acute ischemic stroke due to large vessel occlusion patients who achieved successful recanalization (modified Thrombolysis in Cerebral Infarction ≥ 2b) and underwent TCD monitoring within 72 hours after EVT. Model development and performance assessment were conducted using a bootstrap-based internal validation framework with 1000 stratified bootstrap replicates. Within each replicate, preprocessing and least absolute shrinkage and selection operator-penalized logistic regression were performed using the in-bag data for feature selection. Five classifiers (logistic regression, support vector machine, k-nearest neighbors, random forest, and gradient boosting machine) were trained on in-bag samples and evaluated on out-of-bag samples. Performance was assessed using the area under the receiver operating characteristic curve and threshold-based metrics, with calibration and decision curve analyses. Model interpretability was examined using SHapley Additive exPlanations.

Across bootstrap replicates, smoking status, peak systolic velocity (PSV)_Ratio, age, and mean flow velocity (MFV)_Index were the most frequently selected predictors. Under internal validation, all models showed stable discrimination, and the random forest demonstrated the most favorable overall performance, with good calibration and net benefit. SHapley Additive exPlanations identified MFV_Index as the most influential feature, followed by PSV_Ratio, smoking status, and age. Higher MFV_Index, higher PSV_Ratio, smoking, and older age were associated with a higher predicted risk of unfavorable outcomes.

Using bootstrap-based internal validation, we developed an interpretable machine learning model integrating post-EVT TCD hemodynamics and clinical factors that showed promising internally validated performance for predicting 3-month functional outcomes after EVT. MFV_Index, PSV_Ratio, smoking status, and age were key predictors, highlighting the potential relevance of postreperfusion cerebral hemodynamics and baseline vascular risk.