The trade-off between maximizing reconstruction and physiological interpretation of muscle synergies with autoencoders

TL;DR

This paper explores how different autoencoder designs affect the extraction of muscle synergies, finding that best reconstruction doesn't always mean best physiological interpretation.

Contribution

The study identifies that ReLU+tanh autoencoder configurations yield more physiologically meaningful synergies than those with best reconstruction accuracy.

Findings

Extracted synergies are highly sensitive to autoencoder architecture choices.

ReLU+tanh configurations produce more physiologically meaningful synergies than those with optimal reconstruction accuracy.

Non-linear techniques are emphasized for better synergy extraction across various movement datasets.

Abstract

In neuroscience, the muscle synergy method is a widely known computational approach for studying motor control from electromyographic (EMG) recordings. Standard algorithms for synergy extraction rely on a linearity assumption for synergy combination. However, the interactions between muscle groups and movement dynamics often exhibit non-linear characteristics, suggesting the need for alternative approaches. In this context, autoencoders (AEs) have been proposed as promising tools. However, previous studies focused on the reconstruction accuracy optimization and not on the structure of the synergies, and the influence of AE design parameters has not been thoroughly investigated. This study aims to explore the impact of different activation functions on the effectiveness of AEs.

To this end, we used a rich dataset of upper-limb EMG signals recorded from 16 muscles in 15 participants performing reaching movements toward 9 targets across 5 planes. We evaluated the effects of combining four activation functions in the encoder and decoder layers—linear, ReLU, sigmoid, and tanh—and compared to standard non-negative matrix factorization (NMF).

Our findings show that the extracted synergies are highly sensitive to the AE architecture. Notably, the configurations obtaining the best signal reconstruction do not correspond to the most physiologically meaningful synergies, which were instead achieved with the ReLU+tanh configuration.

This suggests that optimizing reconstruction accuracy may result in non-interpretable synergy structures. This research emphasizes the role of non-linear techniques in extracting muscle synergy from different datasets (e.g., lower limbs, full-body movements, patient populations) and identifies the optimal combination of transfer functions for the encoder and decoder layers.

Figures