Machine learning models based on location-radiomics enable the accurate prediction of early neurological function deterioration for acute stroke in elderly patients

Xiao Ming Fu ¹ Chuan Yang Zhang ¹ Hong Jie Huang ² Chang Cheng Li ² Miao Miao Li ¹ Xiao Ran Li ¹ Zhi Jun Gao ¹ Ming Yang Peng ² Hui Xu ^2* Wen Li Zhu ¹

• ¹ The Affiliated Gaochun Hospital of Jiangsu University, Nanjing, China
• ² Department of Radiology, Nanjing First Hospital, Nanjing Medical University, Nanjing, China

The timely and accurate identification of elderly stroke patients at risk of early neurological deterioration (END) is crucial for guiding clinical management.The present study aimed to create a comprehensive map of lesion location in elderly stroke, and build a machine learning model integrating location features and radiomics to predict END in elderly stroke patients.Methods: A cohort of 709 elderly stroke patients from two centers patients were enrolled. Three machine learning models (logistic regression (LR), random forest (RF), and support vector machine (SVM)) based on location features, radiomics, and Loc-Rad were constructed to predict END in elderly stroke patients, respectively. The performance of models was evaluated using the receiver operating characteristic curves 删除了: L 删除了: models (ROC) and decision curve analysis (DCA). The SHapley Additive exPlanations (SHAP) was used to interpret and visualize the impact of the model predictors on the risk of END. Results: The location maps for elderly stroke patients showed the bilateral cerebellum, left basal ganglia, left corona radiata, and right occipital lobe were significantly associated with END (P < 0.05). For three ML algorithms, the Loc-Rad models based on location features and radiomics demonstrated better performance than the separate location and radiomics -based models in the training cohort (P < 0.05), and the Loc-Rad model constructed with the RF algorithm performed best, with an AUC of 0.883 and accuracy of 0.888, and showed strong prediction performance in the external validation set (AUC of 0.818; accuracy of 0.811). The SHAP plots demonstrated that the most significant contributors to model performance were related to postcentral gyrus left, superior frontal gyrus right, w-HLH_glcm_Correlation, large vessel occlusion and lateral ventricle_body left. Conclusion: We constructed comprehensive maps of strategic lesion network localizations for predicting END in elderly stroke patients and developed a predictive ML model that incorporates both location and radiomics features. This model could facilitate the rapid and robust prediction of the risk of END, enabling timely interventions and personalized treatment strategies to improve patient outcomes.