Abstracts

Rationale: Idiopathic generalized epilepsy (IGE) is classified into subtypes based on clinical manifestations, with juvenile myoclonic epilepsy (JME) and epilepsy with generalized tonic-clonic seizures alone (GTCA) being the predominant subtypes in adults. This study evaluates the performance of five different machine learning models in predicting two types of epilepsy: JME and GTCA. The models utilize the cortical thickness of the brain as a feature.

Methods: The dataset comprised measurements of LH and RH cortical thickness for different brain regions, with each patient labeled as having either JME or GTCA. We utilized the cerebral cortical thickness measurements of 124 JME patients and 36 GTCA patients, encompassing 34 different brain regions. The data were split into training and testing sets in an 8:2 ratio. Five machine learning models were applied: Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), and Gradient Boosting. The features were standardized prior to model training. Model performance was assessed using accuracy, precision, recall, F1-score, area under the ROC curve (AUROC), and the 95% confidence interval (CI) for AUROC, determined through bootstrapping. Additionally, permutation importance was used to determine the top 3 brain cortical regions influencing the Logistic Regression model.

Results: For the 8:2 split, the performance metrics were as follows: Logistic Regression achieved an AUROC of 0.516 (95% CI: 0.479-0.557), Decision Tree 0.531 (95% CI: 0.497-0.564), Random Forest 0.518 (95% CI: 0.477-0.558), SVM 0.484 (95% CI: 0.444-0.528), and Gradient Boosting 0.498 (95% CI: 0.455-0.541). Although all models demonstrated similar overall performance, they generally exhibited low recall values, indicating difficulty in accurately predicting GTCA. Permutation importance analysis for the Logistic Regression model identified the top 3 important features as RH_insula (Importance: 0.0028, Std: 0.0117), RH_paracentral (Importance: 0.0236, Std: 0.0227), and RH_lingual (Importance: 0.0313, Std: 0.0129).


Conclusions: This study highlights the potential of using cortical thickness measurements to classify epilepsy types through machine learning models. Despite the similar performance across models, the low recall values suggest challenges in accurately predicting GTCA. The permutation importance analysis of the Logistic Regression model underscored the significance of specific brain regions, particularly RH_insula, RH_paracentral, and RH_lingual, in influencing predictions.


Funding: None