An exploratory AI-based framework for functional outcome prediction in acute ischemic stroke using DWI, SWI, and clinical metadata

Accurate prediction of functional outcomes in acute ischemic stroke (AIS) is essential for personalized clinical management. However, single-source approaches may be insufficient to capture the combined effects of ischemic lesion characteristics, venous imaging features, and clinical status, while manual delineation of DWI lesions and SWI-derived representative DMV marker is time-consuming and susceptible to inter-observer variability. Therefore, this study aims to develop and evaluate an exploratory AI-based multimodal framework that integrates DWI lesion features, SWI-derived representative DMV marker features, and clinical metadata to predict dichotomized 90-day modified Rankin Scale (mRS) outcomes. This retrospective study included 150 AIS patients for model development and 26 additional patients for external validation. The proposed framework used a diffusion model-based segmentation network to delineate DWI lesions and SWI-derived representative deep medullary vein (DMV) markers. To improve data diversity, a Generative Adversarial Network (GAN)-based augmentation strategy was applied to generate additional lesion and venous marker representations. Image-derived features extracted from the segmented ROIs were then integrated with clinical metadata in a multimodal classification model. A Random Forest classifier and Dempster-Shafer theory were used to combine clinical and imaging information for dichotomized 90-day functional outcome prediction. The proposed segmentation framework achieved a Dice score of 80.05% and an HD95 of 8.14 mm for DWI lesion segmentation, and a Dice score of 40.66% and an HD95 of 10.59 mm for SWI-based representative DMV marker segmentation. For functional outcome prediction, the combined use of DWI, SWI, and clinical metadata achieved the highest accuracy of 93.3% on the internal dataset, with an accuracy of 88.5% on the external validation set. This study presents an exploratory AI-based framework that integrates DWI lesion features, SWI-derived representative DMV marker features, and clinical metadata to predict dichotomized 90-day functional outcomes in AIS. The experimental results demonstrate the feasibility of combining multimodal MRI features with clinical variables for prognosis assessment and provide a methodological basis for future AIS outcome prediction studies. Further prospective, multicenter validation with larger cohorts, calibration assessment, and more comprehensive clinical variables will help strengthen the generalizability and potential clinical applicability of this framework.