Decoding Finger Movements from Medium-Density EMG Using GA-Optimized CNNs and Biomechanical Validation

Accurate interpretation of motor intent is crucial for controlling digital interfaces. Surface electromyography (EMG) is a promising non-invasive solution but faces a trade-off between high-resolution research systems and practical commercial devices. This study presents an integrated approach using a Medium-Density EMG (MD-EMG) system to decode proportional finger movements. We combine wearable hardware, a deep learning decoder optimized with a genetic algorithm, and a biomechanical validation framework to achieve robust and anatomically plausible real-time control. A wearable MD-EMG system with 64 dry electrodes recorded forearm muscle activity. A Convolutional Neural Network (CNN), automatically optimized with a Genetic Algorithm (GA), decoded six degrees of freedom (DoF) of finger movement from raw EMG. Decoded movements were validated for biomechanical plausibility using the MyoSuite simulation environment. System performance was evaluated in offline and real-time tasks with human participants. The GA-optimized CNN achieved high, consistent decoding performance (average offline R² ≈ 0.8). Real-time control performance matched offline results, showing a seamless transition to interactive use. Thumb opposition was the most challenging movement to decode. MyoSuite integration successfully smoothed the decoder's output and ensured movements were biomechanically plausible. Our findings validate that an MD-EMG system, paired with a GA-optimized deep learning model and biomechanical validation, provides accurate, naturalistic decoding of finger movements. This integrated system offers a practical balance between perforamnce and usability, making it a promising solution for applications from virtual reality to advanced prosthetics