Novel gait phases recognition framework leveraging the temporal structure of the myoelectric activity

<i>Objective</i>. Reliable control of lower limb prostheses during gait using surface electromyography requires robust decoding of myoelectric signals to ensure safety and efficiency. Conventional myoelectric pattern recognition (PR) methods, which classify features extracted from each window, often yield inaccurate and unstable output, limiting their practical use.<i>Approach</i>. To deal with these issues, two novel temporal myoelectric-based gait phase recognition frameworks are presented. Temporal activation profile (TAP) considers a sequence of features extracted from consecutive windows, and dual activation shots (DAS) using features extracted from the current and a specific preceding window. These methods were tested on (1) publicly available SIAT-LLMD dataset of 40 healthy subjects under different locomotion conditions, and (2) two subjects with transfemoral amputation during normal walking.<i>Main results</i>. TAP and DAS significantly outperformed conventional PR methods, achieving accuracies of 88.50% and 87.97%, respectively, in healthy subjects during normal walking. TAP achieved optimal performance using features extracted from consecutive windows spanning 240 ms in the past, whereas DAS performed best when leveraging features from the current window combined with those from a window 160 ms prior. No significant differences were observed between TAP and DAS under optimal conditions. Both approaches effectively enhanced gait phase recognition performance when applied to transfemoral amputee gait data. The TAP framework achieved the highest performance, surpassing 87.80% accuracy with extended temporal context requirement, and outperforming the DAS approach (82.32%) under pathological conditions.<i>Significance</i>. Both TAP and DAS are robust solutions for gait phase recognition as they stabilize the decision output and reduce classification errors. DAS is more practically feasible due to lower temporal and computational demands, while TAP is more effective in the case of altered neuromuscular activation patterns. The findings of this paper highlight the potential of integrating these methods into real-time prosthetic controllers, ensuring safe and reliable use for patients.