Deep learning assisted EEG signal processing for controlling bionic hands

Background With the convergence of neuroscience, robotics, and artificial intelligence, several systems have been developed that enable direct brain-machine interfaces. In this context, we designed and implemented a soft robotic hand controlled by brain signals recorded via electroencephalography (EEG). Materials and methods The software system employed deep learning, specifically the ShallowFBCSPNet model, to classify EEG signals corresponding to finger movements. The model was trained under intra-subject conditions using a cropped training method and evaluated across seven sessions. On the hardware side, a soft robotic hand was fabricated using 3D printing technology, actuated by servomotors, and integrated with a hardware-in-the-loop playback interface implemented via Python and Arduino. Results The results demonstrated that training accuracy increased from 84% in the first session to 98% in the seventh session, while validation accuracy improved from 76% to 90%, indicating effective learning and good generalization with minimal overfitting. Error rates for both training and validation sets steadily decreased, and perturbation matrix analysis showed that the model correctly recognized most finger movements, with minor errors mainly occurring between anatomically adjacent fingers. Evaluation metrics, including precision, recall, and F1 score, further confirmed balanced and high performance across all classes. Conclusion Experimental results showed that the system can translate mental commands into robotic hand movements, demonstrating strong potential for motor rehabilitation, human-robot interaction, and other BCI applications.