A multi-stage multimodal learning model for functional outcomes prediction in acute ischemic stroke

Predicting functional outcomes in acute ischemic stroke (AIS) patients after endovascular thrombectomy (EVT) remains challenging due to the complex and dynamic nature of disease progression. Existing methods typically rely on single-stage imaging or limited multimodal fusion, failing to fully capture disease progression and cross-modal complementarity. To address this issue, we propose a multimodal deep learning framework that integrates multi-stage imaging and clinical notes for 90-day modified Rankin Scale (mRS) prediction. The proposed model leverages a shared foundation model with Low-Rank Adaptation to encode pre- and post-treatment images. It further introduces a shared feature extraction module and a stage-specific feature extraction module to disentangle shared and stage-specific representations, enabling effective modeling of disease progression across stages. In addition, a bidirectional cross-attention module is employed to enhance semantic alignment between imaging features and clinical notes. Experiments on multi-stage NCCT and MRI datasets demonstrate that the proposed method consistently outperforms single-modality baselines, achieving AUCs of 0.9147 and 0.9479, respectively. Ablation studies further validate the effectiveness of both the multi-stage interaction design and the multimodal fusion strategy. These results suggest that integrating pre- and post-treatment imaging with clinical notes can improve prognostic performance, providing a practical solution for individualized outcome prediction in AIS patients following EVT.