Temporal lobe epilepsy (TLE) affects millions globally, with a significant portion resistant to drug therapy. Surgery, specifically anterior temporal lobectomy (ATL), is effective for drug-resistant TLE, yet long-term seizure freedom isn't guaranteed for all patients. This suggests the existence of TLE subtypes with varied pathophysiology and treatment responses. Current clinical classifications, relying on radiological or electrophysiological findings, may not adequately predict treatment outcomes. Therefore, the study aimed to identify TLE subtypes using a data-driven approach and non-invasive data (MRI) to improve treatment stratification and prognosis. Magnetic resonance imaging (MRI) is crucial in TLE diagnosis and treatment planning. Studies show a link between seizure frequency, hippocampal atrophy progression, and cortical thinning. However, limited research has explored spatiotemporal patterns of these pathological processes. Recent advancements in machine learning, particularly the Subtype and Stage Inference (SuStaln) algorithm, enable identification of distinct disease progression patterns from cross-sectional data. This algorithm has been successfully applied to other neurological disorders, showing its ability to uncover subtypes based on biomarker changes. The study hypothesized that TLE exhibits phenotypic heterogeneity in atrophy progression, and that SuStaln could identify distinct TLE subtypes with different clinical and treatment characteristics.

The literature highlights the limitations of existing clinical classifications of TLE in predicting treatment outcomes. While imaging studies have shown progressive atrophy in TLE, the spatiotemporal patterns of this atrophy and its relation to subtypes remain unclear. Existing studies have explored the role of hippocampal atrophy and cortical thinning in TLE progression, but a comprehensive understanding of the different pathways of atrophy and their clinical implications was lacking. The SuStaln algorithm's success in other neurological diseases provided a promising methodological approach for this study.

The study used a large cohort of 296 TLE patients and 91 healthy controls. High-resolution T1-weighted MRIs were acquired and processed using FreeSurfer to obtain gray matter volume (GMV) and cortical thickness (CT) measurements for various regions of interest (ROIs). The SuStaln algorithm was then applied to the z-score transformed data of 23 selected ROIs known to be involved in TLE pathology. SuStaln identified distinct trajectories of atrophy progression, and patients were classified into subtypes based on their probability of belonging to each trajectory. The optimal number of clusters (trajectories) was determined using cross-validation. Neuroanatomical signatures of each subtype were visualized using brain mapping. Clinical characteristics (age of onset, illness duration, seizure lateralization, etc.) and long-term treatment outcomes (medication vs. surgery) were compared between subtypes using appropriate statistical tests. Finally, a machine learning model (SVM) was developed to predict surgical outcomes, leveraging SuStaln subtype information and baseline clinical data. The model’s performance was evaluated using ten-fold cross-validation and compared to a model without SuStaln subtype information.

SuStaln identified three distinct trajectories of atrophy progression in TLE: 1) Left hippocampus-predominant, 2) Right hippocampus-predominant, and 3) Cortex-predominant. These trajectories, along with a fourth subtype characterized by amygdala enlargement without atrophy, resulted in four distinct TLE subtypes. Subtypes 1 and 2 (hippocampus-predominant) showed high rates of hippocampal sclerosis (HS) on MRI and ipsilateral seizure lateralization. Subtype 3 (cortex-predominant) had a lower rate of HS and more diffuse cortical atrophy. Subtype 4 ('normal' signature) showed amygdala enlargement and lacked significant atrophy elsewhere. Subtypes differed significantly in clinical characteristics such as age of onset, illness duration, seizure frequency, and aura. Treatment outcomes varied considerably among subtypes. Subtype 4 showed a better response to medication alone, whereas the other subtypes benefited significantly more from anterior temporal lobectomy (ATL) surgery. A subtype-specific machine learning model showed better prediction performance on classifying surgical outcome (seizure freedom) compared to models using only clinical information. The reproducibility of SuStaln subtypes was validated in an independent cohort of 109 TLE patients.

This study provides compelling evidence for phenotypic heterogeneity in TLE. The identification of four distinct subtypes using a data-driven approach improves our understanding of the complex pathophysiological mechanisms underlying TLE. The different atrophy patterns suggest distinct disease progression pathways and potential targets for therapeutic interventions. The findings highlight the clinical relevance of these subtypes, as treatment responses vary significantly. The better prediction of surgical outcome using a subtype-specific machine learning model demonstrates the potential for personalized treatment strategies. The study's findings emphasize the importance of considering brain imaging-based subtyping for improved patient selection and treatment optimization. Further research should investigate the underlying biological mechanisms driving these subtypes and the potential of biomarker-driven personalized medicine in TLE.

This study demonstrates the existence of four distinct biotypes in TLE using machine learning analysis of brain imaging data. These subtypes differ in atrophy patterns, clinical characteristics, and treatment responses. This research highlights the importance of considering these subtypes for improved patient care and personalized medicine. Future studies should focus on larger cohorts, longitudinal follow-up, and exploration of underlying biological mechanisms to further refine TLE subtyping and optimize therapeutic strategies.

While the study employed a large sample size and an independent validation cohort, the reliance on cross-sectional MRI data limits the ability to definitively confirm the temporal progression of atrophy. Further longitudinal studies are needed to track disease progression and validate the proposed subtypes over time. Also, the specific underlying mechanisms driving the differences in subtypes remain to be fully elucidated. More research investigating other biomarkers like brain connectivity and gene expression are necessary for a more comprehensive understanding of TLE heterogeneity.