Estimation of impulse response between electromyogram signals for use in conduction delay distribution estimation.

The time delay between two surface electromyograms (EMGs) acquired along the conduction path is used to estimate mean action potential conduction velocity. Modeling the linear impulse response between "upstream" and "downstream" EMG signals permits an estimate of the distribution of velocities, providing more information. In this work, we analyzed EMG from bipolar electrodes placed on the tibialis anterior of 36 subjects, using an inter-electrode distance of 10 mm. Regularized least squares was used to fit the coefficients of a finite impulse response model. We trained the model on one recording, then tested on two others. The optimum correlation between the model-predicted and actual EMG averaged 0.70. We also compared estimation of the mean conduction delay from the peak time of the impulse response to the "gold standard" peak time of the cross-correlation between the upstream and downstream EMG signals. Optimal models differed from the gold standard by 0.02 ms, on average. Model performance was influenced by the regularization parameters. The impulse responses, however, incorrectly contained substantive power at very low time delays, causing delay distribution estimates to exhibit high probabilities at very short conduction delays. Unrealistic distribution estimates resulted. Larger inter-electrode spacing may be required to alleviate this limitation.

Modeling the impulse response between pairs of EMG signals to estimate conduction delay distribution.

Mean electromyogram (EMG) conduction delay is often estimated as the average time delay between two surface EMG recordings arranged along the conduction path. It has previously been shown that the complete distribution of conduction delays can be estimated from the impulse response relating the "upstream" EMG recording to the "downstream" recording. In this work, we examined regularized least squares methods for estimating the impulse response, namely the pseudo-inverse with small singular values discarded and post hoc lowpass filtering. Performance was evaluated by training the model to one recording, then testing on others. Correlation between model-predicted EMG and measured EMG was assessed for 36 subjects, using EMG recordings with 5 mm inter-electrode spacing. The best correlation was 0.86, on average, for both regularization methods. We additionally compared the mean conduction delay computed from the "gold standard" cross-correlation method to the peak time of the impulse response. The best models differed by 0.01 ms, on average, for both regularization methods. Nonetheless, the impulse responses exhibited excessive energy near zero time, causing delay distribution estimates to exhibit high probabilities at unphysiological short time delays. Inter-electrode spacing larger than 5 mm may be required to alleviate this limitation.