Fusion of Dwi Image and Clinical Variables for Stroke Outcome Prediction Using Tabular Transformer

Ischemic stroke is a major cause of global disability, making outcome prediction crucial for effective management. In this study, we propose an approach using a tabular transformer to fuse diffusion weighted MRI and clinical variables for stroke outcome prediction. Radiomics are employed to extract quantitative features from regions of interest drawn on early diffusion weighted MRI acquired at acute stage in intensive care unit. The 3-month modified Rankin Scale (mRS) score serves as the ground truth for forecasting patient outcomes. Performance of the proposed model was evaluated on a private database encompassing 119 exams of stroke patients treated with thrombectomy (66% of which had a favorable prognosis). The proposed fusion approach achieves strong predictive performance, with an AUC of 0.75 and an accuracy of 0.72. Furthermore, analyzing feature importance contributed to a deeper understanding of clinical relevance.