Automatic Detection of Stimulation Artifacts to Isolate Volitional from Evoked EMG Activity

Control systems for human physiotherapy exercises based on functional electrical stimulation (FES) have provided excellent performance in several setups. Myocontrolled neuro-prostheses use electromyography (EMG) for timing and intensity control of stimulation applied to these exercises, estimating not only the volitional activity (from the patient) but also the evoked activity (from FES). A typical EMG response to FES starts with the stimulation artifact, followed by an excitation curve called M-wave. To extract volitional and evoked components, we first need to find the inter-pulse-intervals (IPIs), i.e., the EMG signal between stimulation artifacts. For that, we have developed a method for two-channel stimulation artifact detection for EMG signals which are not hardware-synchronized to a FES stimulator. First, the artifact detection approach marks all potential artifacts based on one of three adaptive threshold-based detection methods (mean/standard deviation, median/MAD and quantiles). Subsequently, for IPI extraction we cluster the potential stimulation artifacts to cross-correlate the resulting potential stimulation artifact vector with a vector of expected artifacts based on the stimulation and EMG frequencies. For evaluation, we performed tests on two benchmark datasets obtained from FES-assisted walking with two hardware setups. We found more than 95% success rate for both hardware setups using the adaptive threshold method independently on the selected method for choosing the threshold. Because of its low computational demands, we recommend the mean/standard deviation approach.