[Machine learning-based classification of temporal lobe epilepsy subtypes and surgical prognosis evaluation using PET...

Area of Science:

• Neuroscience
• Medical Imaging
• Computational Biology

Context:

• Temporal lobe epilepsy (TLE) is a common neurological disorder characterized by recurrent seizures originating in the temporal lobe.
• Accurate TLE subtype classification and prediction of surgical prognosis are crucial for effective patient management.
• Current diagnostic and prognostic methods may benefit from advanced computational approaches.

Purpose:

• To develop and validate machine learning (ML) models for classifying TLE subtypes and predicting surgical prognosis.
• To identify key brain metabolic network features from ¹⁸F-FDG PET data that contribute to classification and prediction accuracy.
• To compare the performance of various ML algorithms in this task and select the most clinically practical model.

Summary:

• Retrospective analysis of ¹⁸F-FDG PET data from 137 drug-resistant TLE patients (training cohort) and 92 patients (independent test cohort).
• Brain metabolic network connectivity was analyzed using Kullback-Leibler divergence similarity estimation (KLSE), generating 6,902 network attributes.
• Eight ML models, including random forest, were trained for TLE subtype classification and surgical prognosis prediction, with the random forest model demonstrating superior performance (AUC 0.985 in training, 0.946 in testing).
• The models showed high accuracy in predicting surgical outcomes for both mesial TLE (AUC up to 0.838) and neocortical TLE (AUC up to 0.962).

Impact:

• Provides a robust, data-driven approach for TLE subtyping and surgical outcome prediction.
• Enhances clinical decision-making by offering accurate prognostic information for TLE patients.
• Demonstrates the utility of ¹⁸F-FDG PET-derived metabolic network features in conjunction with ML for epilepsy research and clinical application.