Mil Med. 2026 Aug 1;191(Supplement_1):209-215. doi: 10.1093/milmed/usag058.
ABSTRACT
INTRODUCTION: Although powered orthotic devices are becoming more available and increasingly recognized by the medical community, several factors still prevent them from being widely adopted. These orthoses typically use a simple controller, often termed a "bang-bang" controller, which actuates a motor if the muscle's electromyography (EMG) signal is above a set threshold. A proportional controller that could actuate the elbow mechanism based on the EMG signal magnitude would enable a better human-machine connection. The real challenge for designing a proportional controller lies in the heterogeneity of the types of injury and surgeries used in this patient population, as well as the extent of recovery post-surgery. Neural networks effectively model nonlinear relationships in control system design. Recently, a novel powered myoelectric elbow orthosis (PMEO) was developed and tested on patients with a brachial plexus injury (BPI) that used a bang-bang controller. The report evaluated the possibility of using a proportional controller using a neural network-powered regression (NNR) for this exoskeleton.
MATERIALS AND METHODS: Data from 31 participants with a BPI were collected after receiving IRB approval. A custom apparatus was designed to measure elbow flexion torque and EMG, while participants were instructed to match their torque to a predefined target (10%-40% of MVC). A 2-layered NNR (2 hidden nodes) with a Tanh activation and 5 cross-validation folds was used to find a relationship between the properties of the EMG (moving average, root mean square, variance, standard deviation, and slope) and the elbow torque for individual participants. The collected data were divided into a training set (70%) and a test set (30%). The root mean squared error (RMSE) for the test set was calculated for each NNR model, along with the training time and prediction speed.
RESULTS: A standard computer (Intel Core i5-10500 with 16 GB RAM) was used to train the models. The trained models required 82 seconds (range: 3-332 seconds) to train on 8.2e4 (range: 7.7e4-8.9e4) samples. The test dataset consisted of 3.5e4 (range: 3.3e4-3.8e4) samples. The median test RMSE was 0.39 Nm (range: 0.1-3.09 Nm). The NNR took 7 milliseconds (range: 6-10 milliseconds) to predict the test dataset; 35% and 16% of the models had an R2 > 0.5 and >0.7, respectively.
CONCLUSIONS: A low-powered office computer was sufficient to train an NNR with training time short enough to be completed in an orthotist's office within a typical appointment time. These data verify that the process can be field deployable.
PMID:42560192 | DOI:10.1093/milmed/usag058