Developing machine learning models for personalized game-based stroke rehabilitation therapy in virtual reality

Stroke rehabilitation is crucial for restoring patients’ physical and cognitive functions. Traditional rehabilitation methods often lack precision, adaptability, and efficiency, leading to suboptimal recovery. Existing machine learning models, particularly Bi-LSTM, have demonstrated potential in rehabilitation but suffer from scalability issues, imprecise time allocation, and limited explainability. Prior models fail to incorporate optimization techniques, limiting their ability to adapt to dynamic rehabilitation data. They do not adequately leverage performance metrics like completion and error rates, resulting in generalized, rather than personalized, therapy plans. This study introduces an optimized Bidirectional Long Short-Term Memory (Bi-LSTM) model enhanced with Firefly Optimization (FFO). The FFO-Bi-LSTM model refines model parameters using FFO, improving accuracy, convergence speed, and adaptability while minimizing computational overhead. The model is trained using real-world rehabilitation data, optimizing hyper parameters with FFO to enhance efficiency. Key performance metrics such as rehabilitation progress, prediction accuracy, and computational efficiency are analysed. The FFO-Bi-LSTM achieves 99.06 % accuracy, reducing rehabilitation time with personalized therapy plans. Average task completion time is 50 s, with a 98 % success rate and 95 % performance score. Computational efficiency is demonstrated with training and optimization costs of 570 s and an inference time of 0.25 s per sample. Future work will integrate real-time patient monitoring.