Systematic Evaluation of sEMG Processing Pipelines for Gesture Recognition

Abstract

Gesture recognition enables intuitive human–robot communication, where surface electromyography (sEMG) provides a minimally invasive interface for detecting motor activity. However, the lack of systematic evaluation of processing pipelines represents a critical barrier to reliable subject-independent deployment. This work presents a systematic evaluation of sEMG processing pipelines for three-gesture recognition (Neutral, Ask, Take) using a Bagged Trees classifier under Leave-One-Subject-Out (LOSO) cross-validation across ten participants. A mixed-effects ANOVA over 1872 experimental configurations revealed that the preprocessing pipeline is the dominant factor affecting classification accuracy, followed by inter-subject variability, while window size and overlap exhibit smaller but statistically significant effects. A consistency-based elbow analysis identified a compact subset of five features that reduced input dimensionality by 94.79% while improving accuracy from 68.10% to 69.86% and reducing training time by 51.62%. Bayesian hyperparameter optimization was also tested, but it did not yield statistically significant improvements over the five-feature baseline; given the limited cohort (n = 10), this indicates the absence of a detectable difference and points to inter-subject variability as a leading factor limiting performance rather than model configuration. These findings provide empirically grounded guidelines for designing computationally efficient sEMG-based gesture recognition systems for collaborative robotics.