Mapping Motor Cortex Stimulation to Muscle Responses: A Deep Neural Network Modeling Approach

Md Navid Akbar, Mathew Yarossi, Marc Martinez-Gost, Marc A. Sommer, Moritz Dannhauer, Sumientra Rampersad, Dana H. Brooks, Eugene Tunik, Deniz Erdogmus

Introduction

Transcranial Magnetic Stimulation (TMS) is a cornerstone of non-invasive brain stimulation, used to probe and modulate the motor cortex. By inducing magnetic fields, TMS triggers Motor Evoked Potentials (MEPs) in muscles, which are recorded via electromyography (EMG). However, the relationship between the parameters of stimulation—such as coil position and intensity—and the resulting coordinated multi-muscle activation is highly non-linear.

TMS-induced electric field (E-field) modeling in the motor cortex.

Accurately mapping these cortical stimulations to muscle responses is essential for developing targeted therapies for motor impairments following stroke or in conditions like Parkinson’s disease. This paper introduces M2M-Net, a hybrid deep neural network designed to bridge this gap by incorporating biological priors into a regression framework.

Methodology: The M2M-Net Pipeline

The researchers developed a multi-stage forward model to predict muscle activation from TMS parameters. The process begins with Finite Element (FE) Modeling, which transforms TMS coil data into a 3D volumetric electric field (E-field) map of the subject's brain.

Biological Priors and Neural Firing

A key innovation of this work is the conversion of raw E-fields into simulated neural firing rates. Using a non-linear sigmoidal transformation based on primate data, the model estimates the activity of neural ensembles. This step acts as a biological filter, focusing the network on regions most likely to contribute to motor output.

Deep Learning Architecture

The core of the system is a hybrid convolutional neural network (CNN). As shown in the figure below, a 3D convolutional autoencoder first compresses the high-dimensional input into a lower-dimensional latent space.

The M2M-Net pipeline, featuring neural firing conversion and a convolutional autoencoder.

Following compression, a CNN Mapper predicts muscle activation. The authors tested various mapper depths and three output strategies:

  1. Direct: Mapping directly to 15 individual muscle MEPs.
  2. Synergy: Mapping to coordinated muscle groups (synergies) extracted via non-negative matrix factorization.
  3. Both: A hybrid approach mapping to synergies first, then modeling residual errors directly.

Results and Key Insights

The model was trained on 629 TMS trials from a single healthy subject, with intensities ranging from 110% to 140% of the resting motor threshold.

  • The Power of Priors: Converting raw E-fields to neural firing rates significantly improved accuracy. The optimal model achieved a Normalized Root Mean Squared Error (NRMSE) of 0.500 ± 0.043, compared to 0.661 when using raw E-fields.
  • Synergy-Based Modeling: The "Both" output strategy (Synergy + Residuals) proved most effective. This suggests that the network benefits from mimicking the physiological structure of the corticospinal tract, which employs both broad synergistic and fine-grained direct connections.
  • Model Complexity: Statistical evaluations using BIC and AICc confirmed that a shallower 3-layer mapper was optimal when paired with the pre-processed neural firing inputs. Deeper models tended to overfit when provided with these high-quality features.

Conclusion and Future Work

M2M-Net demonstrates that embedding biological knowledge into deep learning architectures can significantly enhance the modeling of complex brain-muscle interactions. While currently limited to a single-subject dataset and exhibiting some variance in error due to low-activation trials, this approach provides a foundation for personalized neurological rehabilitation and more precise diagnostic tools in clinical settings.

Summary updated

BibTeX
@inproceedings{akbar2020mapping,
  author = {Md Navid Akbar and Mathew Yarossi and Marc Martinez-Gost and Marc A. Sommer and Moritz Dannhauer and Sumientra Rampersad and Dana H. Brooks and Eugene Tunik and Deniz Erdogmus},
  title = {{Mapping Motor Cortex Stimulation to Muscle Responses: A Deep Neural Network Modeling Approach}},
  booktitle = {PETRA},
  year = {2020},
  doi = {10.1145/3389189.3389203},
}
Figure