Muscle activity detection from myoelectric signals based on the AR-GARCH model

Myoelectric (EMG) signals contain temporal muscle activation information, that is essential in understanding and diagnosing neuromuscular disorders. Given the biological stochasticity and measurement noise, statistical signal processing methods are adopted in the literature to detect the muscle activity onset and offset periods. However, these methods carry an implicit assumption of stationarity. In this paper, we show that the EMG signal is non-stationary and the nature of its non-stationarity is reminiscent of the heteroscedasticity, i.e., the conditional variance of the signal is time-varying. We therefore model the EMG signal using an Autoregressive-Generalized Autoregressive Conditional Heteroscedastic (AR-GARCH) process, which captures the heteroscedasticity of the signal. The Akaike information criterion test confirms that the AR-GARCH model better fits the EMG signal than the stationary AR model. We subsequently propose a muscle activity detector that relies on the estimated conditional variance of the AR-GARCH model. The application of the proposed detector to real EMG signal shows that the proposed AR-GARCH-based detector achieves a higher accuracy than the widely used double threshold detector.