Decoding Gait Dynamics from Muscle-Specific EMG Using Machine Learning

Optic flow is a fundamental perceptual cue for selfmotion control, closely linked to postural control and neuromuscular activation. This work examines whether visual motion information-stimulus type and optic flow velocity-can be decoded from bilateral, muscle-resolved EMG and quantifies the relative contribution of individual muscles. Surface EMG was recorded bilaterally from four muscles (paraspinal C4, trapezius descendens, tibialis anterior, soleus) in 20 healthy participants exposed to four visual conditions (Baseline, Fovea, Full-field, Peripheral) and five velocities (8, 11, 14, 17, 20 °/s). Two complementary models were evaluated independently: (i) a feature-based Random Forest trained on statistical/spectral descriptors and (ii) a sequence model (Residual CNN-BiGRU with Attention) trained on raw EMG. A Recursive Feature Elimination (RFE) ablation on the Random Forest assessed muscle-level relevance. Both approaches yielded accurate decoding for stimulus identification and velocity estimation, with bilateral and multi-muscle configurations generally outperforming single-muscle and unilateral inputs. The results support EMG-based machine learning (ML) as a practical probe of visuomotor integration and provide a basis for applications in gait recognition, postural assessment, and human-machine interaction.