To evaluate determinants of first-pass reperfusion and to develop an explainable machine learning framework for intra-procedural decision support in patients undergoing endovascular thrombectomy.

In this retrospective single-center study, 204 consecutive patients with acute ischemic stroke treated with endovascular thrombectomy between June 2021 and June 2025 were included. FPE was defined as achieving mTICI 2c–3 reperfusion after a single device pass. A total of 19 clinical, imaging, and procedural variables were analyzed. Six ML models were developed within a structured preprocessing–feature selection–classification pipeline.

Model performance was assessed using repeated nested cross-validation. Discrimination, calibration, and clinical utility were evaluated, and model interpretability was explored using SHAP and partial dependence analyses. The models were designed to reflect real-world intra-procedural conditions.

FPE was achieved in 94/204 patients (46.08%). Favorable collateral status, lower systolic blood pressure, absence of hypertension, lower NIHSS, and higher ASPECTS were significantly associated with FPE (all p < 0.05). Among ML models, Random Forest demonstrated the best overall classification profile, with the highest accuracy, specificity, F1 score, and Youden index, whereas AdaBoost achieved the highest ROC-AUC and PR-AUC values. Explainability analyses consistently identified collateral circulation, systolic blood pressure, hypertension, and age as the most influential predictors. A non-linear relationship between systolic blood pressure and FPE was observed.

Explainable machine learning provided an internally validated framework for estimating the likelihood of first-pass reperfusion using routinely available variables. However, given the limited number of events and absence of external validation, these findings should be considered preliminary and require confirmation in larger independent multicenter cohorts before clinical implementation.