(35) Multi-posture Training Improves EMG-controlled Prosthetic Hand Decoder Performance

Introduction: Over one-third of upper-limb prostheses are rejected by people living with an amputation [1], in part due to limited functionality, including insufficient finger and wrist control [2]. Indeed, most upper limb prostheses lack an actuated wrist, despite its importance in enhancing function. A key challenge in implementing wrist control is understanding its impact on myoelectric decoders. Wrist posture affects the biomechanics of the extrinsic finger muscles, with electromyographic (EMG) activity varying up to 70% for the same finger movement across different wrist postures [3]. This variability could pose a significant challenge as decoders are often trained using EMG data from a single wrist posture. Here, we evaluate how wrist posture affects the classification of finger movements to inform the development of more robust myoelectric decoders.

Methods: Data from a prior study were used, which consisted of intramuscular EMG signals from the forearm muscles of 11 able-bodied people and 3 people with transradial amputations. Participants performed 10 repetitions of individual finger flexion-extension movements across five wrist postures: neutral, flexed, extended, pronated, and supinated. To quantify EMG activity similarity between the neutral posture and the other four postures, the Cross Projection Similarity Index (CPSI) was used. CPSI employs Principal Component Analysis (PCA) to define manifold spaces and measures variance preservation when projecting between configurations. Bidirectional variance ratios were averaged, ranging from 0 (dissimilar) to 1 (identical), and repeated across all wrist configurations. Additionally, a Linear Discriminant Analysis classifier was trained to predict which finger was moving using EMG data (70:30% training - testing data split) with 10-fold cross-validation across all participants.

Results: The EMG data obtained from each individual wrist posture variant (flexed, extended, pronated, supinated) did not fully capture the variance present in the neutral configuration, reducing the CPSI (Figure.1, p < 0.001, Wilcoxon rank-sum test with Bonferroni correction). An LDA classifier achieved 89% accuracy when trained and tested on neutral posture data. However, accuracy dropped to 34–37% when tested on data from other wrist postures. Training the classifier on data from all wrist postures improved accuracy to 68%, outperforming cross-posture classification but remaining below single-posture performance.

Conclusion: Wrist posture significantly impacts EMG activity patterns, as reflected in the reduced CPSI values and decoder performance. Training on diverse postures improves generalization but does not fully resolve accuracy loss. Future work should explore advanced machine learning techniques to model non-linear and temporal dependencies.