Predicting the Path: How Machine Learning Can Identify Subtypes of Epilepsy and Predict Disease Progression

Abstract

Identification of Four Biotypes in Temporal Lobe Epilepsy via Machine Learning on Brain Images

Jiang Y, Li W, Li J, Li X, Zhang H, Sima X, Li L, Wang K, Li Q, Fang J, Jin L, Gong Q, Yao D, Zhou D, Luo C, An D. Nat Commun. 2024;15(1):2221. PMCID: PMC10933450

Artificial intelligence provides an opportunity to try to redefine disease subtypes based on similar pathobiology. Using a machine-learning algorithm (Subtype and Stage Inference) with cross-sectional magnetic resonance imaging from 296 individuals with focal epilepsy originating from the temporal lobe epilepsy (TLE) and 91 healthy controls, we show phenotypic heterogeneity in the pathophysiological progression of TLE. This study was registered in the Chinese Clinical Trials Registry (number: ChiCTR2200062562). We identify 2 hippocampus-predominant phenotypes, characterized by atrophy beginning in the left or right hippocampus; a third cortex-predominant phenotype, characterized by hippocampus atrophy after the neocortex; and a fourth phenotype without atrophy but amygdala enlargement. These 4 subtypes are replicated in the independent validation cohort (109 individuals). These subtypes show differences in neuroanatomical signature, disease progression, and epilepsy characteristics. Five-year follow-up observations of these individuals reveal differential seizure outcomes among subtypes, indicating that specific subtypes may benefit from temporal surgery or pharmacological treatment. These findings suggest a diverse pathobiological basis underlying focal epilepsy that potentially yields stratification and prognostication—a necessary step for precise medicine.

Commentary

Patient heterogeneity in epilepsy has long been observed and gives rise to difficulties in effectively diagnosing and treating the disorder. However, recent work has begun to more closely examine patterns of heterogeneity, revealing the presence of subtypes within epileptic populations. In temporal lobe epilepsy (TLE), 3 to 4 cognitive phenotypes have been identified,^1,2 and it has been suggested that a new taxonomy of phenotypes based on cognitive and behavioral classifications could be useful in the development of precision medicine treatments. ³ Of course, the question remains how specific types of structural changes in the brain might give rise to these clinically observable distinct phenotypes. A recent study by Jiang et al ⁴ provides important insight into this question by applying a recently developed machine learning algorithm called Subtype and Stage Inference (SuStaIn) to magnetic resonance imaging (MRI) images from TLE patients in order to identify 4 distinct spatiotemporal trajectories of neural atrophy.

While the application of the SuStaIn algorithm to data from TLE patients is novel, its application to MRI in neurodegenerative diseases is not. The SuStaIn algorithm was originally developed to identify the presence of distinct trajectories of neurodegeneration in dementia and Alzheimer's patients, revealing subtypes within these communities that could be linked to known genotypes. ⁵ It was further employed on imaging data from patients with schizophrenia, revealing 2 distinct trajectories of brain atrophy. ⁶ In order to understand the success of the SuStaIn algorithm, it helps to understand a bit about how the algorithm works. Importantly, the SuStaIn algorithm does not require longitudinal data, despite its ability to identify temporal trajectories. Instead, the method relies on the fact that in large cross-sectional data sets involving many patients, each individual is likely to be at a different point in their disease progression when the imaging is performed. The user of the algorithm selects a number of biomarkers (generally regional gray matter volumes) and these are measured from the MRI image of each patient. This set of biomarkers for each patient is used as the input to the algorithm. The algorithm then assumes a piecewise linear progression from the beginning state of a given biomarker to its end state—for neurodegenerative diseases, this corresponds to an initial regional brain volume to a final (smaller) regional brain volume. Using a sophisticated mathematical model, the algorithm optimizes the number of distinct trajectories and builds spatiotemporal trajectories from the snapshots of biomarker data across patients. For those interested in a more detailed description of the exact models and methods used, the original publication ⁵ has an excellent explanation of the method.

Jiang et al ⁴ capitalized on the known presence of neural atrophy in TLE to apply this methodology to a large cross-sectional dataset of patients with TLE. Importantly, this dataset also involved 2 different treatment groups (medical and surgical) such that subtypes identified through patterns of brain atrophy progression could also be linked to clinical outcomes. The study identified 3 distinct spatiotemporal patterns of atrophy that were characterized as trajectory 1: atrophy starting in the left hippocampus; trajectory 2: atrophy starting in the right hippocampus; and trajectory 3: atrophy starting in cortical regions. Each trajectory additionally expressed a specific spatial pattern of atrophy spreading throughout the brain over time. The algorithm also identified another subgroup of patients that did not display any atrophy relative to control patients, resulting in a final identification of 4 subtypes of TLE patients.

Patients were relatively well distributed across the 4 subtypes, with 28.7%, 38.2%, 13.9%, and 19.2% classified as belonging to the left hippocampal, right hippocampal, cortex, and “normal” subtypes, respectively. As expected, the first 2 hippocampal subtypes were more likely to have hippocampal sclerosis and display seizure lateralization in the corresponding hemisphere. It was also observed that the cortical subtype had a significantly higher total intercranial volume than the other 3 subgroups.

Of more interest was how patients in the different subtypes differentially responded to different treatments (medication or anterior temporal lobe surgery). Patients in the “normal” subtype who were provided with medication-based treatment were significantly more likely to report seizure freedom than patients assigned to the other 3 subgroups who received medication-based treatment. However, patients in the “normal” subtype who received surgery-based treatment were significantly less likely than other subtypes to report seizure freedom at follow-up. In fact, patients in the first 2 hippocampal subgroups responded relatively well to surgical treatment, with 69.2% and 72.2% reporting seizure freedom, respectively.

Finally, the authors showed that the prediction of surgical outcomes could be improved if a patient's subtype was considered. More specifically, the authors assumed that the proper prediction for surgical outcomes might depend on different features for patients of each subtype and built 4 subclassifiers, 1 for each subtype identified by the SuStaIn algorithm. These subclassifiers performed reasonably well at predicting surgical outcomes, showing a statistically significant improvement over random predictions. Importantly, each subclassifier also performed better than a classifier built on a data set that did not account for subtypes, and a classifier trained using clinical baseline information did not perform better than random predictions.

While the fact that the SuStaIn algorithm uses cross-sectional data to identify distinct brain trajectories is a useful feature, it remains important to validate the identified trajectories on longitudinal datasets. Unfortunately, large, longitudinal neuroimaging datasets of epileptic patients are not currently available. Future studies should therefore focus on filling this important gap in data.

Perhaps the most exciting aspect of this work is the emphasis on shifting paradigms to emphasize the use of data-driven approaches to identify subtypes within epilepsy. While the current study only looks at the effects of regional brain atrophy to identify brain trajectories associated with TLE, it is known that epilepsy is associated with brain network reorganization. It will therefore be interesting to develop similar data-driven methods that can identify distinct patterns of changes in brain network structure: brain network trajectories. Initial work to identify subtypes (but not associated trajectories) based on network properties has been performed on structural morphological networks^7,8 and functional networks ⁷ of epileptic patients. Other work studying changes in tractography networks of rodents posttraumatic brain injury (post-TBI) similarly found subtypes of distinct patterns of network reorganization post-TBI. Based on simulated brain activity, the authors speculated that 1 subtype might be more likely to develop epilepsy. ⁹ Like the SuStaIn study, none of these analyses were performed on longitudinal datasets, again emphasizing the need to collect this important data.

The ability to identify specific types and trajectories of brain network changes that can be associated with different mechanisms of epileptogenesis, disease severity, and/or clinical outcomes will be essential in developing personalized treatment strategies, and the work by Jiang et al ⁴ provides an important step forward in achieving that goal.