Pattern Recognition Control System for a Four-Electrode Prosthetic Hand Using a Spatio-Temporal Feature-Based CNN

Upper limb amputation significantly impacts individuals’ autonomy and quality of life, making the development of advanced prosthetic devices a key area of biomedical research. In recent years, myoelectric prostheses—driven by surface electromyographic (sEMG) signals—have gained increasing attention from researchers and companies, showing great promise in restoring motor function through intuitive control systems. This work presents a pattern recognition-based approach for controlling a four-electrode myoelectric hand prosthesis using a spatio-temporal feature-based Convolutional Neural Network (CNN). To support the development and validation of the proposed method, a dedicated sEMG database was created, referred to as the "EH Database", named after the prototype device “Electrode Hub” provided by BionIT Labs. The database comprises multi-day recordings from ten limb-intact participants performing eight distinct hand and wrist gestures across three different upper limb postures. An acquisition protocol was designed to introduce intra-subject variability and ensure high-quality signal segmentation, enabling training and validation of the control model. Subject-specific five-class models were trained using the full EH Database for the selected movements. Classification accuracies above 95% were achieved in 9 out of 10 subjects on Test Set 1, which included data acquired on the same day as the training data. Furthermore, performance remained high on Test Set 2, which was composed of data acquired during an additional session, simulating prosthesis use on a different day from the calibration phase; in this case, 7 out of 10 subjects achieved accuracies exceeding 90%. To simulate the prosthesis calibration phase, additional models were trained using a reduced version of the Training Set, corresponding to only nine seconds of acquisition per class. Despite the limited amount of training data, these models achieved accuracies above 95% for 9 out of 10 subjects. On Test Set 2, accuracy slightly decreased but still exceeded 90% in 5 out of 10 subjects. A comparison was conducted with a four-channel version of the current control system implemented on the Adam’s Hand prosthesis from BionIT Labs. The proposed CNN model consistently outperformed the aforementioned approach, with average margins exceeding 20% on Test Set 1 and slightly under 15% on Test Set 2. The results of this work demonstrate strong model performance in an offline context, suggesting the potential for future testing in a real-time application. Moreover, the study highlights the importance of adding incremental updates to the training dataset through subsequent calibration sessions, which can significantly enhance the model’s robustness across different usage days.