Identification of Subject-Specific EMG Detectors for Robot-Assisted Therapy Using Rest and Move State EMG Data

In severely impaired stroke subjects without visible movements, implementing robot-assisted therapy based on movement intention decoded from electromyogram (EMG) requires both sufficient residual EMG to drive robotic assistance and a subject-specific detector to ensure accurate, low-latency detection. However, identifying such a detector is challenging, particularly when the presence of residual EMG in a given subject is unknown. This paper proposes a systematic approach to distinguish between EMG data when the subject is relaxed versus attempting a movement. We investigated six different detector types and separation measures using retrospective EMG data from a previous randomized controlled trial. The results indicate that the approximate generalized likelihood ratio (AGLR) detector, along with the modified Hodges and modified Lidierth detectors, achieved the best separation between the data when a subject is relaxed compared to when he/she attempts movements. Using a subset of clinician-annotated data to evaluate detection performance, the modified Hodges detector combined with probability difference-sum ratio measure (MH-PDSR) showed good performance in terms of both accuracy and latency. Based on the EMG data from 30 severe stroke, we propose a PDSR threshold of 0.7 with the modified Hodges detector to identify stroke subjects with sufficient residual EMG. These findings suggest that the MH-PDSR approach can be used to learn a maximally separating detector for a given subject which can be used both to screen stroke subjects for residual EMG and to provide a detector to drive robotic assistance if residual EMG is present. Further validation using larger datasets and evaluation of the resulting human machine interaction is warranted.