Stroke rehabilitation remains a critical challenge in neurorehabilitation due to its reliance on timely assessment, individualized therapy, and continuous monitoring. This study proposes an AI-driven, sensor-based framework that leverages multimodal data to enhance stroke recovery prediction. The objective is to deliver an accurate, interpretable, and scalable solution that integrates brain, movement, and physiological signals for real-time decision support. A novel Stacked Deep Ensemble Learning (SDEL) architecture is introduced, comprising specialized base learners (CNN-LSTM, EEG Transformer, and vitals-based DNN) and a Gradient Boosting meta-learner. Multimodal sensor data were collected, preprocessed using advanced denoising and feature extraction techniques, and fused for outcome modeling. Experimental results show the SDEL model outperforms baseline architectures, achieving 94% accuracy, 0.96 AUC-ROC, and robust generalization on unseen data. The model’s interpretability via SHAP analysis and its resilience to partial modality failures make it a strong candidate for clinical deployment. This work lays the foundation for adaptive, AI-powered rehabilitation platforms.