Decomposing MRI phenotypic heterogeneity in epilepsy: a step towards personalized classification

Area of Science:

• Neurology
• Neuroimaging
• Machine Learning

Background:

• Drug-resistant temporal lobe epilepsy (TLE) presents challenges in predicting individual patient outcomes.
• Current group-level analyses overlook critical inter-patient heterogeneity, hindering personalized care.
• Understanding individual variations is key to advancing person-centered epilepsy treatment.

Purpose of the Study:

• To apply unsupervised machine learning to identify underlying disease factors in TLE.
• To analyze inter-individual variability in structural brain pathology using multimodal MRI.
• To improve predictions of clinical outcomes and cognitive dysfunction in TLE patients.

Main Methods:

• Utilized unsupervised machine learning on 3T multimodal MRI data (cortical thickness, hippocampal volume, FLAIR, T1/FLAIR, diffusion parameters) from 82 TLE patients.
• Estimated four latent disease factors representing whole-brain structural pathology patterns.
• Validated factor specificity against healthy controls and frontal lobe epilepsy patients.

Main Results:

• Identified four robust and specific latent disease factors characterized by distinct patterns of atrophy, gliosis, and microstructural alterations.
• Classifiers trained on these factors predicted drug response (76% ± 3%) and postsurgical seizure outcome (88% ± 2%) with high accuracy.
• Latent factor models significantly outperformed baseline learners in predicting inter-patient cognitive dysfunction.

Conclusions:

• Data-driven identification of latent disease factors offers a novel approach to understanding TLE heterogeneity.
• These factors capture complex, interacting pathological processes contributing to individual disease variability.
• Incorporating inter-individual variability analysis holds promise for enhancing clinical prognostics and personalized TLE management.