sEMG‐Based Explainable Neural Networks Using Transfer Learning for Intersubject Finger‐Joint‐Angle Estimation

Finger‐joint‐angle (FJA) estimation based on surface electromyographic (sEMG) signals plays an important role in the control of prosthetics and exoskeletons. However, most of the existing FJA‐estimation methods are unexplainable and subject‐specific, and accurate FJA estimation remains a challenge. This study pioneeringly modeled the relationship between the forearm muscles and hand movements for FJA estimation by adding a multihead self‐attention (MHSA) block between long short‐term memory networks. Additionally, an intersubject transfer‐learning method based on the K‐nearest‐neighbors clustering‐based pretraining strategy is proposed to ensure similarity between the data used for pretraining and that of a new user. The model performance is evaluated on the Ninapro DB2 dataset and collected data. Results show that the proposed method significantly outperformed the state‐of‐the‐art methods in terms of root‐mean‐square error (Ninapro DB2: 6.37 ± 0.16 vs. 7.26 ± 0.18, p < 0.01; collected data: 5.64 ± 0.21 vs. 5.08 ± 0.15, p < 0.01) and correlation coefficient (Ninapro DB2: 0.87 ± 0.01 vs. 0.85 ± 0.01, p < 0.01; collected data: 0.92 ± 0.01 vs. 0.88 ± 0.01, p < 0.01). Moreover, the proposed model is based on an explainable prediction mechanism wherein the importance of information from the different muscles is quantified by the MHSA.