Abstract
BACKGROUND: Malignant cerebral edema (MCE) is a severe complication of acute ischemic stroke (AIS) that is associated with poor outcomes or death. The study sought to develop a predictive machine learning (ML)-based model for MCE following AIS using radiomics features from non-contrast computed tomography images of the infarct lesion (IL), the affected hemisphere (AH), and the whole brain (WB). METHODS: A total of 219 AIS patients from four centers were included in this study. Patients from Centers 1, 2, and 3 were allocated to a training cohort and a test cohort by stratified randomization at a ratio of 8:2, while those from Center 4 were allocated to an independent external validation cohort. Radiomics features of the IL, the AH, and the WB were extracted. After the feature selection process, the radiomics features related to MCE were identified. Using seven distinct ML algorithms, an IL model based solely on IL radiomics features, and a combined IWA model that incorporated IL, AH, and WB radiomics features were developed. The performance of the models were assessed by calculating the area under the curve (AUC) value. RESULTS: The IWA model demonstrated effectiveness in predicting MCE risk, with the multilayer perceptron-based model achieving particularly high performance. The IWA model had a higher AUC than the IL model (0.927 vs. 0.865, P<0.05). CONCLUSIONS: This study developed a novel IWA model that was able to effectively predict the risk of MCE following AIS and was superior to the IL model. It is expected that our model will provide more precise guidance recommendations for clinical treatment in the future.