A low-cost transhumeral prosthesis operated via an ML-assisted EEG-head gesture control system.

Key challenges in upper limb prosthetics include a lack of effective control systems, the often invasive surgical requirements of brain-controlled limbs, and prohibitive costs. As a result, disuse rates remain high despite potential for increased quality of life. To address these concerns, this project developed a low cost, noninvasive transhumeral neuroprosthesis-operated via a combination of electroencephalography (EEG) signals and head gestures.To address the shortcomings of current noninvasive neural monitoring techniques-namely, single-channel EEG-we leveraged machine learning (ML), creating a neural network-based EEG interpretation algorithm. ML generation was guided by two underlying goals: (1) to improve overall system performance by combining discrete models using a prediction voting scheme, and (2) to favor modelwithin these new neural network ensembles, as opposed to individual model. EEG data from eight frequency bands was collected from human subjects to train a ML algorithm employing a hierarchical mixture-of-experts structure. We also implemented head gesture-based control to assist in the generation of additional stable classes for the control system.The final model performs competitively with existing EEG interpretation systems. Inertial measurement unit (IMU)-based head gestures supplement the neural control system, with 270° actuation of synovial elbow and radial wrist joints driven by intuitive corresponding head gestures. The brain-controlled prosthesis presented in this study costs US$300 to manufacture and achieved competitive performance on a Box and Block Test.These results suggest proof-of-concept for potential application as an alternative to current prosthetics, but it is important to note that the demonstration in this study remains exploratory. Future work includes broader clinical testing and exploring further uses for the developed ML system.