GaitNet: Transfer Learning-Enhanced CNN-GRU Architecture for Intention Detection in Healthy and Post-Stroke...

Highly accurate, data-efficient, and real-time detection of human motion intention is essential for the effective control of assistive and rehabilitation robots in post-stroke recovery. However, existing approaches focused on healthy participants and often require large amounts of subject-specific data, limiting their clinical applicability. To address these challenges, we propose GaitNet, a novel transfer learning-enhanced dual branch CNN-GRU framework designed to deliver robust and data-efficient motion intention detection for both healthy and post-stroke participants. GaitNet combines convolutional layers for spatial feature learning, cross-attention for latent space fusion, and gated recurrent units (GRUs) for temporal modelling, while leveraging transfer learning to reduce reliance on extensive subject-specific data. In evaluations, using full amount of data, GaitNet achieved an accuracy of 99.29%±0.17 on healthy participants and 98.96%±0.47 on post-stroke participants, outperforming baseline deep learning and classical models. In transfer learning, GaitNet was pre-trained on healthy participants and fine-tuned on post-stroke data. Remarkably, under joint supervised pretraining (JSP), it maintained > 98% accuracy, precision, recall, and F1-score with only 45% of the training set on post-stroke participants (9 repetitions per motion), reducing the data requirement by 65% while preserving high performance. Furthermore, with an inference time of 27.46 ms per input batch, GaitNet operates well within the real-time constraints reported in related assistive robotics studies and was successfully deployed in a closed-loop robotic control setting with healthy participants, confirming its suitability for real-time rehabilitation scenarios. These results establish GaitNet as a practical and scalable solution for highly accurate, data-efficient detection of motion intention in post-stroke participants.