Abstracts

Rationale: Juvenile myoclonic epilepsy (JME) is a prevalent form of epilepsy with heterogeneous prognoses influenced by both clinical and radiological factors. Although the clinical characteristics of JME have been extensively studied, the integration of these multimodal data into robust predictive models remains limited. Leveraging recent advances in machine learning and neuroimaging, this study aimed to develop and externally validate artificial intelligence models that integrate clinical and radiological features for prognostic prediction in JME patients.

Methods: We conducted a retrospective study including 125 patients diagnosed with JME. Comprehensive clinical data, including demographic information, seizure history, and treatment details. were collected. Structural MRI data were analyzed using volumetric and cortical thickness measurements, as well as advanced radiomics features. Multiple machine learning models (logistic regression, random forest, extreme gradient boosting [XGBoost], light gradient boosting machine [LightGBM], support vector machine [SVM], and artificial neural network [ANN]) were developed. Performance was evaluated with accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUROC). Models were trained on an internal dataset and independently validated on an external cohort.

Results: Key predictors of favorable prognosis included male gender, left amygdala and right hippocampal volumes, and cortical thickness in the bilateral temporal poles, left entorhinal cortex, fusiform gyrus, and right inferior/middle temporal cortex. Integrated models combining clinical, volumetric, cortical thickness, and radiomics data outperformed models relying on a single modality. Among them, the random forest model achieved the best performance, with an AUROC of 0.923. The multimodal approach proved especially powerful, highlighting the role of brain structures such as the thalamus and hippocampus, consistent with current insights into JME pathophysiology.

Conclusions: This study underscores the value of up-to-date machine learning approaches that integrate clinical and radiological data to enhance prognostic prediction in JME. Multimodal data-driven models demonstrated superior predictive accuracy compared to single-source models, reflecting a promising trend in precision epilepsy care. Incorporating advanced neuroimaging features alongside clinical variables may facilitate more informed treatment strategies and improve outcome prediction in clinical practice. Future studies with larger, more diverse populations and additional imaging modalities, such as diffusion tensor imaging and functional MRI, will be essential for further validation and refinement of these models.

Funding: None