Recognizing EEG responses to active TMS vs. sham stimulations in different TMS-EEG datasets: a machine learning...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Transcranial Magnetic Stimulation (TMS) combined with Electroencephalogram (TMS-EEG) assesses cortical neuron properties.
• TMS-evoked EEG potentials (TEPs) can include non-neuronal artifacts, necessitating quality control tools.

Purpose of the Study:

• To differentiate EEG responses to TMS from sham stimulations using machine learning.
• To evaluate the accuracy of a machine learning model in identifying true TMS effects.

Main Methods:

• Utilized two independent TMS-EEG datasets from healthy volunteers (N=33).
• Employed a Bi-directional Long Short-Term Memory (BiLSTM) network to classify EEG signals from TMS and sham conditions.
• Assessed accuracy at single-trial and multi-trial (5-20 trials) levels.

Main Results:

• The BiLSTM model achieved moderate to high accuracy (60-75% single-trial, >75% for 20 trials) in distinguishing TMS from sham conditions.
• Sham conditions generally showed lower accuracy, except for unmasked auditory stimulation.
• Comparisons between pre-stimulus TMS and sham EEG yielded chance-level accuracy, while post-stimulus comparisons showed moderate to high accuracy.

Conclusions:

• TMS-evoked potentials are distinguishable from sham stimulations using a BiLSTM machine learning approach.
• The model demonstrates effectiveness even at the single-subject level and with limited trials.
• This provides an accurate, automatic tool for assessing TEP quality.