On autoencoders for extracting muscle synergies: A study in highly variable upper limb movements

The muscle synergy method is a well-established computational approach to motor control in neuroscience. Recently, the hypothesis of linearity adopted by conventional algorithms has been questioned since the non-linearities of the musculoskeletal system may not be captured by linear methods. The scope of this work is to shed further light on the capabilities of autoencoders (AEs) for extracting muscle synergies by targeting a variety of movements covering the upper limb workspace. This approach elicits multiple muscular activations, which are essential to exploit the potential of muscle synergies. We developed two configurations of an autoencoder: a single-plane model trained and tested on the same movement plane and a multiple-plane model trained on all planes but tested on each plane. Electromyographic data were collected from 16 muscles of 15 participants performing reaching movements across 9 targets in 5 planes, and results were compared to the non-negative factorization (NMF). Both synergies and temporal coefficients showed high similarity between AE and NMF (>0.78), indicating that the motor modules extracted with the two methods have the same structure and similar temporal recruitment. Both methods showed a comparable reconstruction accuracy of the input signal (RMSE and R2). The performance of AE decreased with multiple plane training with respect to single plane training due to signal variability. Limitations of this study include the lack of ground truth and unexplored AE configurations. To foster future work, we released an open codebase to provide an easy-to-use code for reproducing our study and for testing new features that may improve the application of the AE (https://github.com/cbrambilla/MuscleSynergyExtractionBench-main). Future research will focus on the development of non-linear techniques to extract muscle synergy in different datasets (e.g., lower limbs, full-body movements, patient populations), applying different setting parameters, multi-layer architectures, and activation functions, and incorporating task performance within synergy models.