Incremental Deep Learning for Single-Finger Myoelectric Control: A Comparative Closed-Loop Study of CNN and LSTM Architectures

Kurzfassung

Target achievement control tests provide a standardized way to quantify closed-loop myoelectric control during continuous target acquisition. We compared two electromyography-to-kinematics regression controllers for single-finger control using 32-channel surface electromyography: a pretrained dilated residual convolutional neural network regressor (996k parameters) and a lightweight long shortterm memory regressor without pretraining (65k parameters). Ten non-disabled participants completed a counterbalanced within-subject crossover study. For each controller condition, participant-specific training was performed and target achievement control was evaluated in two blocks: before and after a targeted incremental fine-tuning step using additional data collected from difficult targets. Performance was assessed using completion rate, median time-to-success, overshoot count, and time spent outside the target region after first entry. After participant-specific training, both controllers achieved comparable target achievement control performance across all primary outcomes (|dz|<=0.16, all p>0.6). As the CNN was pretrained on only six participants, this equivalence need not hold at larger pretraining scales. Incremental fine-tuning produced large improvements for both controllers (dz>0.8), increasing completion rate by 12-17 percentage points and reducing movement time by 1.0-1.4 s and post-entry instability by 3-5 s per target. These results suggest that participantspecific incremental learning can enable a lightweight model without pretraining to match a deeper pretrained architecture in single-finger myoelectric control, while targeted fine-tuning provides a robust mechanism to improve speed and stability during use.