A Biophysical-Model-Informed Source Separation Framework For EMG Decomposition

TL;DR

This work introduces BMISS, a Biophysical-Model-Informed Source Separation framework that inverts anatomically accurate forward EMG models to perform unsupervised motor unit decomposition from $sEMG$ and simultaneously infer motor neuron properties. By replacing discrete MUAP libraries with a differentiable MUAP generator anchored in MRI-derived anatomy, BMISS enables gradient-based optimization to recover neural drive and MU properties, while integrating a biophysical foundation through forward modeling like the quasi-static Poisson formulation. In simulated validation, BMISS achieved higher fidelity MU estimation and substantially reduced computation compared to traditional BSS, retrieving the majority of MUs and obtaining MU-property estimates with low $\text{MSE}$ under favorable signal-to-noise conditions; however, performance degrades with noise, highlighting the need for robust real-data validation. The approach offers a path toward non-invasive, personalized neuromuscular assessments with potential clinical, prosthetic, and rehabilitation applications, and suggests future directions for universal, dynamic forward models and hybrid techniques that blend physics-based modeling with data-driven inversion.

Abstract

Recent advances in neural interfacing have enabled significant improvements in human-computer interaction, rehabilitation, and neuromuscular diagnostics. Motor unit (MU) decomposition from surface electromyography (sEMG) is a key technique for extracting neural drive information, but traditional blind source separation (BSS) methods fail to incorporate biophysical constraints, limiting their accuracy and interpretability. In this work, we introduce a novel Biophysical-Model-Informed Source Separation (BMISS) framework, which integrates anatomically accurate forward EMG models into the decomposition process. By leveraging MRI-based anatomical reconstructions and generative modeling, our approach enables direct inversion of a biophysically accurate forward model to estimate both neural drive and motor neuron properties in an unsupervised manner. Empirical validation in a controlled simulated setting demonstrates that BMISS achieves higher fidelity motor unit estimation while significantly reducing computational cost compared to traditional methods. This framework paves the way for non-invasive, personalized neuromuscular assessments, with potential applications in clinical diagnostics, prosthetic control, and neurorehabilitation.

Figures (4)

• Figure 1: Left: Simulated EMG forward model overview from volume conductor MRI to MUAPs to EMG.Right: Instead of performing costly discrete library-based search, instead train a continuous MUAP generative model to interpolate across samples, that enables the use of gradient based methods
• Figure 2: Full decomposition framework. The framework attempts to optimize the input parameters $\bm{c}$ to maximize kurtosis on the separated sources. In the forward pass the model generate MUAPs, to separation vectors to sources. The backward pass involves backpropagating throughout our entire framework.
• Figure 3: 2D slice of the true optimization landscape across different input variables. Heatmaps were generated by evaluating a grid of the loss function (negative kurtosis) over the displayed parameters while holding all other parameters at their true values. The true condition is marked with a red circle. Although the optimization landscape exhibits extended flat regions and highly non-convex features, a distinct local minimum is clearly visible.
• Figure 4: Cumulative number of motor units retrieved per iteration. As observed, using a larger batch size increases the number of points optimized via gradient descent, enabling the retrieval of a greater proportion of the total motor units.