Muscle electrophysiology, particularly surface electromyography (sEMG), is increasingly used in human-machine interfaces (HMIs) beyond traditional clinical applications, extending to robotics and virtual reality. However, current decoding algorithms struggle to meet the fine control demands of these applications. Deep learning offers a promising solution, but its success hinges on access to substantial, high-quality, annotated EMG data. Acquiring such data is costly and time-consuming, as it necessitates recording from multiple subjects under diverse conditions and accurately annotating underlying physiological parameters (e.g., individual muscle forces, fiber properties). Traditional machine learning methods, while suitable for simpler tasks, often require subject-specific calibration and recalibration, hindering their applicability in mass-market HMIs. Data augmentation through simulation is a potential solution, used successfully in other deep learning fields. Existing EMG simulation methods are either not realistic enough or computationally prohibitive, preventing their use in data augmentation for deep learning. This research addresses this gap by introducing a novel EMG simulation method that is both computationally efficient and highly realistic.

Classical uses of biosignals include physiological studies, clinical diagnostics, and monitoring. More recently, their use in HMIs has gained significant traction. Deep learning approaches, coupled with advancements in wearable and affordable recording devices, have further fueled this interest. However, achieving highly functional and intuitive EMG-based HMIs for the mass market requires overcoming challenges related to user-specific calibration and anatomical/physiological variability. Traditional machine learning techniques fall short in this regard. While deep learning holds promise, the need for large, annotated datasets presents a significant hurdle. Existing data augmentation techniques for electrophysiological signals often utilize black-box models that lack physiological relevance, limiting their usefulness. Sophisticated biophysical models exist, but their computational cost is prohibitive for generating the necessary training data. Analytical models based on simplified geometries capture broad signal characteristics but lack the realism for specific experimental conditions. More realistic models based on numerical solutions of the Poisson equation, while accurate, are far too computationally intensive for data augmentation purposes.

The researchers developed a novel EMG simulation method based on the numerical solution of forward equations, tailored for deep learning data augmentation. This method achieves high realism and computational efficiency by exploiting the mathematical and structural properties of the model, reformulating them theoretically. It produces highly realistic EMG recordings while providing access to all underlying physiological parameters. The method's speed surpasses state-of-the-art methods by orders of magnitude (minutes vs. hours), allowing the generation of extensive, high-quality datasets. The Myoelectric Digital Twin is a cloud-based software with a Python API. Users define simulation parameters, including subject anatomy (muscle, bone, fat, skin surfaces), tissue conductivities, electrode configuration, individual fiber properties (neuromuscular junction location, tendon lengths, action potential propagation velocities), motor unit and recruitment model parameters, and muscle activations. The software's architecture, driven by the mathematical properties of the model, permits efficient management of pre-computed data. The core of the method involves a novel approach to solving the forward problem of the volume conductor under electrostatic conditions. This includes a hierarchical decomposition of the EMG simulation pipeline enabling the reuse and optimization of individual steps. Realistic anatomy is discretized into a tetrahedral mesh, and an anisotropic conductivity tensor is associated with each tetrahedron. Instead of solving quasi-static Maxwell's equations for each fiber source and time instant, the researchers solve them for a set of unit point sources at mesh vertices (basis sources), significantly reducing computational time. The adjoint method is employed to further enhance efficiency, solving systems of equations for electrodes rather than basis sources. The model's accuracy was validated by comparing the numerical solution to the analytical solution for a cylindrical volume conductor, demonstrating low error (3-5%). The simulator's performance was evaluated at various scales: single fiber activation in the brachioradialis muscle, excitation of a single muscle (Brachioradialis with 50,000 fibers), and simultaneous excitation of multiple muscles during wrist flexion/extension and abduction. Comparisons with experimental data showed strong qualitative similarities in signal characteristics (time and frequency domains) despite the non-personalized nature of the simulations. The model achieves significant speed improvements without optimization tricks or parallel computing, instead relying on the inherent efficiency of the theoretical reformulation.

The developed Myoelectric Digital Twin is highly computationally efficient, generating EMG signals from anatomically accurate models with tens of thousands of muscle fibers in seconds, a significant improvement over state-of-the-art methods which take hours. The model accurately reproduces various features observed in experimental EMG signals, including propagating and non-propagating components, amplitude increases with muscle excitation, and distinct patterns during flexion, extension, and abduction. Comparisons between simulated and experimental EMG signals using root mean square (RMS) values and spectral analysis demonstrated good agreement, particularly when simulation parameters were adjusted (a simple form of inverse modeling was possible due to the simulation speed). The researchers demonstrated the application of the simulated data for pre-training a deep neural network designed to decompose high-density EMG signals into underlying spinal motor neuron activities. Using the simulated EMG data for pre-training significantly improved the network's performance on experimental data, as measured by the rate of agreement (ROA) metric. The pre-trained network exhibited a significantly higher ROA (93.8% vs. 82.4%) compared to a network with random initialization. This improvement was statistically significant and consistent across both male and female subjects. The improvement was especially prominent for smaller amplitude MUAPs that are difficult to detect in experimental settings.

The Myoelectric Digital Twin provides a highly realistic and computationally efficient approach to surface EMG modeling, enabling the generation of large, varied, and perfectly annotated EMG datasets. This capability opens new avenues for data augmentation in deep learning, which was previously infeasible with existing simulation methods. The achieved computational efficiency significantly surpasses previous efforts by effectively exploiting the mathematical structure of the forward equations and the model architecture. While the model currently omits some sources of variability present in experimental data (advanced noise, biomechanical modeling, non-stationary volume conductor properties), these can be addressed in future improvements. The model also does not automate the estimation of muscle forces for specific movements. Future integration with biomechanical modeling or inverse problem-solving algorithms using synthetically-generated data could address this.

This research introduces a novel, highly efficient EMG simulator, the Myoelectric Digital Twin, capable of generating realistic and personalized EMG data at scale. This tool facilitates data augmentation for deep learning, addressing a major bottleneck in the development of advanced human-machine interfaces. The demonstration of improved performance in spinal motor neuron activity decomposition using pre-training with simulated data underscores the model's potential for transformative impact in the field. Future directions include incorporating additional sources of signal variability and integrating biomechanical modeling to enhance realism and facilitate applications such as personalized HMI development.

The current model does not fully incorporate certain sources of variability observed in real-world EMG signals. These include detailed noise and artifact descriptions, a complete biomechanical model of the musculoskeletal system, and the dynamic nature of volume conductor properties and fiber geometry. Furthermore, automated modeling of muscle forces for specific movements is not currently included, though this is a potential avenue for future development.