Real-time myoelectric control continues to improve beyond eight EMG channels in an LDA-based pattern recognition system

Pattern recognition (PR) is widely used for myoelectric upper limb prosthesis control. PR system performance depends on electromyography (EMG) channel count and extracted feature set, which have historically been selected to minimize classifier error rate. However, offline error rate does not directly reflect user-in-the-loop control, making online (i.e., real-time) evaluations necessary. This study examined the impacts of common channel counts and feature sets on performance of a Linear Discriminant Analysis (LDA)-based PR system using offline, online, and subjective analyses in participants without limb difference (n=12) and with transradial limb loss (n=5). Across feature sets, analysis types, and populations, increasing channel count beyond the conventional limit of eight improved performance. On average, participants without and with limb loss completed a real-time control task 1.79 s and 1.26 s faster (p<0.05), respectively, using 16-channel systems. Feature set expansion produced mixed results but was more beneficial at lower channel counts. Offline metrics moderately predicted online performance (Average \(R^2\) =0.28 and 0.29 for those without and with limb loss, respectively), indicating that error rate alone is informative but insufficient to fully characterize performance without complementary online and subjective evaluations.