Association of Amputation-Related Clinical Variables with Subject-Specific sEMG Hand-Gesture Classification in Transradial Amputees: A Comparative Secondary Analysis of CNN and Convolutional Vision Transformer Models

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Surface electromyography (sEMG)-based hand-gesture recognition in transradial amputees is challenged by substantial inter-subject variability related to physiological, clinical, and experimental factors. This study presents a secondary subject-level analysis of previously validated classification results to examine whether clinical characteristics are associated with performance. Accuracy values from convolutional neural network (CNN) and Convolutional Vision Transformer (CViT) models were matched with clinical metadata from NinaPro DB3 (n = 11) and the MYO amputee cohort (n = 10). A prespecified linear regression framework was complemented by 95% confidence intervals, adjusted R2, Benjamini–Hochberg false discovery rate (FDR) correction, Spearman correlations, residual diagnostics, Cook’s distance, sensitivity analyses, and variance inflation factors. In NinaPro DB3, CNN accuracy showed nominal associations with remaining forearm percentage (β = 0.0044, R2 = 0.509, p = 0.0137) and phantom limb sensation (β = 0.0650, R2 = 0.421, p = 0.0309), but neither remained below the 0.05 threshold after FDR correction and both were strongly influenced by subject A7. No robust univariate clinical association was identified for CViT in NinaPro DB3 or for either model in MYO. These findings suggest that clinical heterogeneity may contribute to performance variability, but the evidence remains exploratory rather than confirmatory.