Fall-related gait pattern recognition based on surface electromyography using a hybrid neural network with transfer learning

Gait analysis across different individuals presents significant challenges due to individual variability. In this study, we developed a gait analysis system to enable continuous decoding of gait patterns based on surface electromyography (EMG). Our objective was to address the long-term challenges associated with neural networks, particularly their limited generalization capability and the heavy training burden on new subjects in gait pattern recognition. Fifteen healthy subjects were recruited to imitate two fall-related gait patterns and collect myoelectric signals from eight lower limb muscles. We proposed a transfer learning-based framework and investigated eight adaptation schemes using varying quantities of target training samples. In the cross-subject evaluation, our framework demonstrated robust generalization capability and reduced training burden. Even with a limited number of target training samples, the framework outperformed state-of-the-art machine learning models and discriminative neural networks that have previously shown excellent recognition performance in subject-specific evaluation. This study highlights the promising potential of using transfer learning in conjunction with neural networks for gait pattern recognition across different individuals. It also provides valuable insights into the diagnosis, monitoring, and treatment of various medical conditions using surface EMG-based gait analysis systems.