Classification of TMS evoked potentials using ERP time signatures and SVM versus deep learning

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biophysics

Background:

• Modeling transcranial magnetic stimulation (TMS) evoked potentials (TEP) is crucial for validating generative dynamical models.
• Current methods often rely on dimensionality reduction techniques to extract features from experimental data.

Purpose of the Study:

• To compare the efficacy of different classification schemes for TEP data.
• To evaluate a deep learning architecture against traditional methods for TEP analysis.

Main Methods:

• Designed a 3-dimensional feature space based on event-related potentials (ERP).
• Compared classification using the 3D feature space versus the original full-rank input space.
• Utilized a deep learning architecture (CNN and MLP) and Support Vector Machines (SVM).

Main Results:

• The deep learning architecture (CNN-MLP) demonstrated superior accuracy in discriminating TEPs.
• This deep learning model outperformed traditional methods using 3D projection or raw TEP input with SVM.

Conclusions:

• Supervised feature extraction models, particularly deep learning, offer enhanced accuracy for TEP classification.
• These models can effectively score neural circuit simulations by assessing their ability to reproduce dynamical processes underlying TEP responses.