Improving knee joint angle prediction through Dynamic Contextual Focus and Gated Linear Units

Real-time, accurate knee joint angle prediction is crucial in biomechanics and rehabilitation, where precision supports improved patient outcomes and more responsive exoskeleton control. This paper introduces FocalGatedNet, a novel deep learning framework combining Dynamic Contextual Focus (DCF) Attention and Gated Linear Units (GLUs) to enhance feature dependency capture, making it highly effective for multi-step gait trajectory prediction. Unlike conventional approaches that rely solely on recurrent or convolutional architectures, FocalGatedNet leverages attention-based mechanisms tailored for time-series forecasting, ensuring superior temporal dependency modeling. Our extensive evaluation of FocalGatedNet on a comprehensive, multimodal gait dataset compares it against top-performing models across multiple prediction intervals (20 ms, 60 ms, 80 ms, and 100 ms). Results show that FocalGatedNet delivers substantial gains in predictive accuracy, with marked improvements in Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). Notably, FocalGatedNet consistently outperforms transformer-based models, demonstrating enhanced robustness across varying movement conditions. For instance, at the 80 ms prediction window, FocalGatedNet achieves reductions in MAE by up to 24%, RMSE by up to 14%, and MAPE by up to 36% over the Transformer model, highlighting its effectiveness in capturing complex knee joint movement patterns. Additionally, we conduct an ablation study to validate the role of GLU and DCF Attention in performance gains, confirming that feature gating significantly enhances model efficiency. Experimental evaluations also assess the impact of sensor noise on prediction accuracy, ensuring real-world applicability. Also, FocalGatedNet works with less time consumption than many other deep learning models. Its efficient inference speed, coupled with high accuracy, makes it a viable solution for deployment in real-time gait analysis and exoskeleton-assisted rehabilitation. Thus, FocalGatedNet is quite helpful and relatively reliable for real-time biomechanical applications. The model implementation is accessible in the GitHub repository: https://github.com/LyesSaadSaoud/FocalGatedNet.