Machine Learning to Predict Impaired Consciousness in Focal and Generalized Epilepsy
Abstract number :
3.299
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2024
Submission ID :
253
Source :
www.aesnet.org
Presentation date :
12/9/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Aya Khalaf, PhD – Yale School of Medicine
Matthew Gruen, BS – Yale University School of Medicine
Erica Johnson, MD – Yale School of Medicine
Basil Abdalla, MD – Yale School of Medicine
Max Springer, BS – University of Maryland
Gregory Worrell, MD, PhD – Mayo Clinic
Lawrence Hirsch, MD – Yale University School of Medicine
Hal Blumenfeld, MD, PhD – Yale University School of Medicine
Rationale: Seizures with impaired consciousness severely affect morbidity, mortality, and quality of life for people with epilepsy. A great challenge arises for clinicians when patients have epileptiform EEG discharges, but there is no available behavioral assessment to determine whether these discharges impair or spare consciousness. We introduce a machine learning approach to predict impaired consciousness in focal and generalized onset seizures based on EEG without the need for behavioral testing. This approach will facilitate medical decision making and allow clinicians to precisely evaluate ability of patients to complete daily-life activities such as driving.
Methods: We developed a machine learning pipeline to predict focal and generalized onset seizures with impaired consciousness based on pre-ictal and ictal EEG data. The algorithm was trained with time and frequency domain features as well as spatial features extracted through applying the common spatial pattern algorithm. Support vector machines and linear discriminant analysis were employed to classify seizures as spared or impaired with the classification performed on pre-ictal and ictal periods separately and combined through feature vector concatenation and probabilistic fusion. The classifier was tested and validated on a data set of 130 seizures in 34 patients with generalized epilepsy, and is being tested further on a larger data set of patients with focal epilepsy currently being collected.
Results: We evaluated classification models with different feature sets, classifiers, and analysis time windows (pre-ictal, ictal, and both combined) on the generalized epilepsy dataset using 10-fold cross validation. The best model achieved a false discovery rate of zero with corresponding sensitivity over 90%. Analysis of behavior in focal seizures has so far shown more variable types of impairment than in generalized seizures. The application of the model and analysis of EEG classifier features in focal seizures is ongoing.
Conclusions: The proposed approach demonstrates the feasibility of EEG-based machine learning systems in clinical practice to predict impaired consciousness associated with focal and generalized onset seizures. Our results showed that time, frequency, and spatial domain features of pre-ictal and ictal EEG can predict seizure impact on consciousness. Successful implementation of this approach in clinical settings will yield a highly useful tool lowering morbidity and mortality of people with epilepsy and improving their quality of life.
Funding: NIH K99NS133494
Neurophysiology