基于脑影像及临床特征的机器学习模型预测缺血性卒中后心房颤动

Objective To investigate the predictive value of a machine learning model based on brain imaging and clinical features in patients with atrial fibrillation detected after stroke. Methods A retrospective cohort design was used in this study. Data was derived from the ischemic stroke and TIA patients enrolled in the China national stroke registry Ⅲ from August 2015 to March 2018. Patients were divided into two groups according to the systematic collection of past medical history records, electrocardiogram, and 24-hour Holter monitoring results during hospitalization: the sinus rhythm group and the atrial fibrillation detected after stroke group. Firstly, a pre-trained nnUNet deep learning framework was applied for standardized preprocessing and automated lesion segmentation of DWI data. Subsequently, 960 quantitative imaging features encompassing eight categories, including morphological characteristics, first-order statistics, and advanced texture features, were extracted using the PyRadiomics open-source package. During the feature engineering stage, the Spearman’s rank correlation coefficient analysis was applied (preset threshold |ρ|>0.8) to eliminate highly collinear features. After retaining independent features, the least absolute shrinkage and selection operator (LASSO) regression algorithm was used for feature selection and to construct a joint prediction model. The model performance was internally validated via five-fold cross-validation, and the AUC of the ROC curve was used as the primary evaluation indicator. Finally, the SHapley Additive exPlanations framework was used to analyze the importance of features. Results A total of 1464 ischemic stroke patients were included, with an average age of (64.5±11.1) years, including 498 patients with atrial fibrillation detected after stroke and 966 patients with sinus rhythm. The average AUC of five-fold cross-validation of the prediction model for atrial fibrillation detected after stroke constructed using 15 clinical features was 0.71 (95%CI 0.67-0.74). Clinical and imaging features were fused to form 975 multimodal features, with an average AUC of 0.73 (95%CI 0.70-0.76). Using the LASSO algorithm for feature selection, 31 multimodal features (including 25 imaging and 6 clinical features) were obtained after screening, with an average AUC of 0.73 (95%CI 0.70-0.77). Conclusions The machine learning model based on brain imaging and clinical features can effectively predict atrial fibrillation detected after stroke, and can be further applied in clinical practice.