This study aims to construct a predictive model for post-thrombectomy hemorrhagic transformation (HT) by integrating hemodynamic features derived from quantitative DSA (qDSA) with machine learning models. A retrospective analysis was conducted on patients with acute anterior circulation large-vessel occlusion who underwent MT at our center from January to December 2024. Immediate postoperative anteroposterior and lateral angiograms of the target vessels were obtained after MT. Hemodynamic features such as mean transit time (MTT) and time to peak (TTP) were all calculated based on the time density curve derived from postoperative DSA. Clinical baseline data and interventional procedural information were also collected. Hemodynamic features associated with post-thrombectomy HT were selected using 5 different feature selection algorithms. Five machine learning models were employed to fit the features influencing the outcomes, and the results were evaluated using the Receiver operating characteristic (ROC) curves. The predictive performance of different models was compared using the area under the curve (AUC). SHapley Additive exPlanations (SHAP) were used for model interpretation. A total of 171 patients with anterior circulation large-vessel occlusion who underwent MT were included. Hemorrhagic transformation occurred in 68 patients. In contrast, the non-HT group comprised patients with the absence of intracranial hemorrhage or large cerebral infarction or cerebral herniation. In 171 patients, 39 hemodynamic parameters within the ROIs of the target vessels were extracted from perfusion images generated by postoperative DSA in each patient. After feature selection, the best-performing model on the testing set was the Elastic-Logistic model consisted of qDSA and clinical features, with an average AUC of 0.86. And the best-performing model on the testing set consisted of qDSA features alone, was the most frequently selected qDSA features still constructed using a logistic model, with an average AUC of 0.81. By integrating clinical and hemodynamic features, machine learning algorithms can be effectively utilized to construct a preliminary predictive model for post-thrombectomy HT in patients with acute ischemic stroke treated by mechanical thrombectomy.
Li et al. (Thu,) studied this question.