Linear vs. non-linear mapping of peak power using surface EMG features during dynamic fatiguing contractions

Abstract

This study compares a non-linear (neural network) and a linear (linear regression) power mapping using a set of features of the surface electromyogram recorded from the vastus medialis and lateralis muscles. Fifteen healthy participants performed 5 sets of 10 repetitions leg press using the individual maximum load corresponding what they could perform 10 times (10RM) with 120 s of rest between them. The following sEMG variables were computed from each extension contraction and used as inputs to both approaches: mean average value (MAV), median frequency (Fmed), the spectral parameter proposed by Dimitrov (FInsm5), average (over the observation interval) of the instantaneous mean frequency obtained from a Choi–Williams distribution (MFM), and wavelet indices ratio between moments at different scales (WIRM1551, WIRM1M51, WIRM1522, WIRE51, and WIRW51). The non-linear mapping (neural network) provided higher correlation coefficients and signal-to-noise ratios values (although not significantly different) between the actual and the estimated changes of power compared to linear mapping (linear regression) using the sEMG variables alone and a combination of WIRW51 and MFM (obtained by a stepwise multiple linear regression). In conclusion, non-linear mapping of force loss during dynamic knee extension exercise showed higher signal-to-noise ratio and correlation coefficients between the actual and estimated power output compared to linear mapping. However, since no significant differences were observed between linear and non-linear approaches, both were equally valid to estimate changes in peak power during fatiguing repetitive leg extension exercise.