A Multimodal Causal Deep Learning Framework for Personalized Stroke Rehabilitation Outcome Prediction and Treatment Recommendation

Stroke causes major long-term disability, with balance impairment significantly affecting quality of life. Personalized prognosis and treatment selection, particularly between TeleRehabilitation (TR) and Conventional Rehabilitation (CR), are crucial. However, current predictive models often lack multimodal integration or tailored recommendations. This paper introduces Causal-MMFNet, a novel deep learning framework. It integrates diverse multimodal time-series data to simultaneously predict balance recovery and allocate individualized treatments in stroke rehabilitation. Key innovations include a dynamic cross-modal attention fusion mechanism, an Individual Treatment Effect (ITE) estimation module for counterfactual outcomes, and causal consistency regularization. Evaluated on the StrokeBalance-Sim dataset, Causal-MMFNet consistently outperforms baselines and state-of-the-art multimodal frameworks, demonstrating superior accuracy and reliability across established metrics. Ablation studies confirm component contributions, while dynamic attention reveals adaptive modality prioritization. The framework's treatment allocation significantly improves patient outcomes, with uncertainty estimation providing clinical confidence. Causal-MMFNet offers a robust, causally-aware solution for personalized decision support in stroke rehabilitation, enhancing patient recovery and optimizing resource allocation.