Log In

Forgot Password?
Create New Account

Loading... please wait

Abstract Details

Using Machine Learning to Augment Differentiation of Epileptic and Psychogenic Non-Epileptic Seizures Recorded with sEMG
Epilepsy/Clinical Neurophysiology (EEG)
P6 - Poster Session 6 (12:00 PM-1:00 PM)
12-001

Continuous sEMG recordings from the biceps muscle have been used in recent studies to differentiate between ES and PNES.  Using empirical observation, features were extracted and manually evaluated for usefulness in classifying events as ES or non-ES. Automated classification of those seizures showed excellent accuracy for labelling TC seizures as ES (100%) but only moderate accuracy for labelling non-TC ES as ES (71%) and PNES as non-ES (79%). This study aimed to improve ML performance in differentiation of non-TC ES from PNES.

To use machine learning (ML) to optimize differentiation of non-tonic-clonic (non-TC) epileptic seizures (ES) from psychogenic non-epileptic seizures (PNES) using surface electromyography (sEMG).

Fifty-five events, including 37 non-TC ES and 18 PNES, were collected during previous studies from 4 EMUs in the US and Austria and retrospectively evaluated. Seizures were verified by three epileptologists’ review of vEEG (majority rules). Twenty-four features were extracted based evidence from discrete and continuous wavelet transforms, local maxima from derivatives, and amplitude-based analyses. The pattern recognition artificial neural network toolbox from MATLAB 2019a was used to identify the most valuable features for ML classification. A single hidden network ML model was used so that the value of each feature could be more easily determined. For training, validation, and testing, 38, 6, and 11 events were used, respectively.

ML model performance accuracy for the training, validation, and testing was 94.7%, 100%, and 81.8% respectively. Of the 24 evaluated input features, 16 features were weighted within an order of magnitude of the highest weighted feature.

ML models trained on sEMG signal data from the biceps muscle may be able to classify non-TC seizure events as  either ES or PNES. Optimized extraction techniques and ML model approaches will be investigated and presented at the conference.

Authors/Disclosures
Damon P. Cardenas, PhD (Brain Sentinel)
PRESENTER
No disclosure on file
Luke Whitmire, PhD (Brain Sentinel) No disclosure on file
Jonathan J. Halford, MD (Medical University of South Carolina) Dr. Halford has received personal compensation in the range of $500-$4,999 for serving as a Consultant for SK Life Sciences. Dr. Halford has received personal compensation in the range of $500-$4,999 for serving as a Consultant for Takeda. Dr. Halford has received stock or an ownership interest from Corticare. The institution of Dr. Halford has received research support from Takeda. The institution of Dr. Halford has received research support from SK Life Sciences. The institution of Dr. Halford has received research support from Biogen.