Machine learning, clinical-radiomics approach with HIM for hemorrhagic transformation prediction after thrombectomy and treatment.

Background: This study aimed to develop a clinical-radiomics model using hyperattenuated imaging markers (HIM), characterized by hyperattenuation on head non-contrast computed tomography immediately after thrombectomy, to predict the risk of hemorrhagic transformation (HT) in patients undergoing endovascular mechanical thrombectomy (MT). Methods: A total of 159 consecutive patients with HIM were screened immediately after MT for inclusion. The datasets were randomly divided into training and test cohorts at a ratio of 8:2. An optimal machine learning (ML) algorithm was used for model development. Subsequently, models for clinical, radiomics, and clinical-radiomics were developed. The performance of the models was measured using receiver operating characteristic (ROC) and decision curve analyses (DCA). The interpretability and predictor importance of the model were analyzed using Shapley additive explanations. Results: Of the 159 patients, 100 (62.9%) exhibited HT. The support vector machine (SVM) was the optimal ML algorithm for constructing the models. In predicting HT, the areas under the curve (AUCs) of the clinical model were 0.918 (95% confidence interval [CI] = 0.869-0.966) in the training cohort and 0.854 (95% CI = 0.724-0.984) in the test cohort. The AUCs of the radiomics model were 0.869 (95% CI = 0.802-0.936) and 0.829 (95% CI = 0.668-0.990), while those of the clinical-radiomics model were 0.944 (95% CI = 0.905-0.984) and 0.925 (95% CI = 0.832-1.000). Conclusion: The suggested clinical-radiomics model based on HIM is a reliable method that can provide a risk evaluation of HT in individuals undergoing MT.