Poster AA10 - Intention Detection for Robotic Acute Stroke Rehabilitation - A Pilot Study

Introduction: Electromyography (EMG) is the process of measuring muscle electrical activity. EMG signals can be used as inputs to a human-machine interface for myoelectric control. Recent developments in rehabilitation robotics use myoelectric control to classify or decode user intent (e.g., hand gestures). Deep neural networks can be particularly useful for this purpose because of the high number of features than can be extracted from EMG signals. Many neural network models are built with or rely on uncompromised EMG signals sourced from healthy individuals. However, the target population for rehabilitation applications are likely to have compromised EMGs. In particular, patients recovering from stroke exhibit zero to non-zero activity EMG in the extremities of the affected side. While non-zero EMG activity can be restored in most instances, movement may not be observed either due to the diminished amplitude of the signal, opposing co-contractions, or other factors. Gesture classification built on top of EMG pattern recognition can be an intuitive control for rehabilitative robotics to assist patients to complete intended gestures. This provides a positive feedback control that reinforces the gesture made by the patient. Gesture classification, or intention detection, is the first step towards gesture decoding, where EMG is processed near real time.

Materials and

Methods: The MindRove Armband, an 8 channel EMG sensor with a 500 Hz sampling rate, was used to record EMG signals from the proximal end of the forearm. Participants were limited to one control and one recovering stroke patient, two months post-stroke. The Hand of Hope (HoH) hand robot, used in a concurrent study was used to allow the participant to perform isometric contractions. The participants were directed to perform a sequence of gestures, or tasks. Both participants performed 30 tasks, each of which lasted for 3 seconds with 6 seconds of rest in between each task. The healthy control performed 6 different types of tasks: OPEN, CLOSE, TRIPOD OPEN, TRIPOD CLOSE, BOTTOM OPEN, and BOTTOM CLOSE. The stroke patient performed 2 tasks: OPEN and CLOSE. Post-processing of the EMG included passing through a bandpass filter to detrend and remove motion artifacts, and performing RMS rectification to convert the signal to the positive value regime. This resulted in a feature set of 8 temporal domain descriptors. A deep neural network with three hidden layers was used to train the model using the processed EMG data. Each layer was 250, 64, and 32 fully connected neurons wide respectively, and applied a ReLU function (slope=0.01). Two training instances were developed, each using 50 epochs. Results were determined based on the performance of the models during training, and were validated against a test set using a classification matrix and 3D t-distributed Stochastic Neighbor Embedding.

Results, Conclusions, and Discussions: The results showed improved performance for both the control and stroke group. Accuracy reached 95% and 89% for the control and stroke group respectively. The loss, which is a quantification of the error of the model, reached 5% and 21% for the control and stroke group. The classification matrix, which compares a set of predicted values against true values provides a quick visual cue of the model performance. The control group had accuracy rates of 0.95, 0.96, 0.96, 0.96, 0.82, and 0.98 for the 6 tasks, open, close, tripod open, tripod close, bottom open, and bottom close. The stroke group had accuracy rates of 0.91 and 0.86 for the two tasks, open and close. T-distributed stochastic neighbor embedding is an unsupervised dimensionality reduction technique for data visualization. It reduced the 8 dimensions of EMG data (i.e., the 8 rectified channels) into 3 dimensions that can be visualized. The control group shows distinct clusters form in the embedded space that corresponds with the 6 tasks. The stroke group shows two clusters, but there is a higher degree of overlap. This region of overlap indicates that these values have higher rates of mislabeling.

These results show that a simple deep fully connected network can classify hand gestures from healthy and compromised EMG signals with an accuracy higher than 85%. The hyperparameters chosen for the model were arbitrarily chosen and not optimized. Many machine learning models are trained using extracted features from a data set. The more features that are extracted from a dataset, the more parameters there are to train on. This increases the performance of the model and its ability to classify incoming data. Tailoring the model for EMG data and performing more feature extraction can increase the performance of the model and optimize it to receive and decode new EMG data. The results from this model can be used to decode EMG signals in a positive feedback control that reinforces patient intention and improves functional outcomes.

Acknowledgements (Optional): I would like to thank Giovanni for his feedback on my methods and results, as well as Dali for providing insights into the mechanisms and anatomy of hand gestures. I would finally like to thank Dr. Zhang for giving me the opportunity to independently conduct research.