ABSTRACT NUMBER: ESOC2026A1025 CONSTRUCTION OF A DEEP LEARNING PREDICTION MODEL FOR THE RECURRENCE RISK OF DIABETIC STROKE BASED ON RADIOMICS

Abstract Background and aims Stroke damages the blood-brain barrier, increasing recurrence risk, an effect worsened by diabetes. Consequently, diabetic stroke patients face higher recurrence and disability rates, severely impacting quality of life. The characteristics of recurrence in this population remain underexplored. This study therefore aims to develop a machine learning-based model to predict recurrence risk in diabetic stroke. Methods A retrospective cohort study enrolled 164 patients with type 2 diabetes and ischemic stroke. Multi-sequence MRI (T1, T2 FLAIR, DWI) data were collected. Based on one-year follow-up, patients were classified into recurrence and non-recurrence groups. Lesions in the basal ganglia and pons were manually segmented as regions of interest. After image preprocessing and Z-score feature standardization, LASSO regression was used for feature selection to reduce dimensionality. Multiple machine learning models, including XGBoost and SVM, were trained and compared. The optimal model was selected based on the highest AUC. Finally, SHAP analysis was applied to interpret the model's predictions globally and locally. Results In the training cohort and the test cohort, the area under the curve (AUC) of the T1-T2FLAIR-DWI sequence fusion model was 0.86 and 0.81, the ACC was 0.77 and 0.76, the sensitivity was 0.72 and 0.64, the specificity was 0.81 and 0.85, and the F1 was 0.72 and 0.70, respectively. Conclusions The machine learning model based on multi-sequence MRI-based radiomics can predict the recurrence risk of diabetic stroke patients and provide data support for individualized secondary prevention. Conflict of interest “Sijia Chen.nothing to disclose”. “Jiawei Liang.nothing to disclose”. “Zhenzhen Wan.nothing to disclose”. “Donghua Mi.nothing to disclose”.