Epoch length and autoregressive-order selection for electromyography signals

This study shows how different EMG-epoch lengths affect the selection of the autoregressive-model orders. Electromyography signals were divided in 25ms, 50ms, 100ms, 250ms and 500ms epochs. Order-selection criteria were applied to the least-square errors of autoregressive models. The Bayesian Information Criterion and the Minimum Description Length indicated that needle-EMG signals recorded from normal subjects at 25kHz could be represented by autoregressive models using orders below 25 for 500ms epochs, and that smaller orders could be used to represent shorter epochs.