Entering the Era of Personalized Medicine in Epilepsy Through Neuroimaging Machine Learning

Decomposing MRI Phenotypic Heterogeneity in Epilepsy: A Step Towards Personalized Classification

Lee HM, Fadaie F, Gill R, et al. Brain 2021;awab425. doi:10.1093/brain/awab425. Online ahead of print.

In drug-resistant temporal lobe epilepsy (TLE), precise predictions of drug response, surgical outcome, and cognitive dysfunction at an individual level remain challenging. A possible explanation may lie in the dominant “one-size-fits-all” group-level analytical approaches that do not allow parsing interindividual variations along the disease spectrum. Conversely, analyzing inter-patient heterogeneity is increasingly recognized as a step toward person-centered care. Here, we utilized unsupervised machine learning to estimate latent relations (or disease factors) from 3T multimodal MRI features (cortical thickness, hippocampal volume, FLAIR, T1/FLAIR, and diffusion parameters) representing whole-brain patterns of structural pathology in 82 TLE patients. We assessed the specificity of our approach against age- and sex-matched healthy individuals and a cohort of frontal lobe epilepsy patients with histologically verified focal cortical dysplasia. We identified four latent disease factors variably co-expressed within each patient and characterized by ipsilateral hippocampal microstructural alterations, loss of myelin and atrophy (Factor-1), bilateral paralimbic and hippocampal gliosis (Factor-2), bilateral neocortical atrophy (Factor-3), and bilateral white matter microstructural alterations (Factor-4). Bootstrap analysis and parameter variations supported high stability and robustness of these factors. Moreover, they were not expressed in healthy controls and only negligibly in disease controls, supporting specificity. Supervised classifiers trained on latent disease factors could predict patient-specific drug response in 76 ± 3% and postsurgical seizure outcome in 88 ± 2%, outperforming classifiers that did not operate on latent factor information. Latent factor models predicted inter-patient variability in cognitive dysfunction (verbal IQ: r = 0.40 ± 0.03; memory: r = 0.35 ± 0.03; sequential motor tapping: r = 0.36 ± 0.04), again outperforming baseline learners. Data-driven analysis of disease factors provides a novel appraisal of the continuum of interindividual variability, which is likely determined by multiple interacting pathological processes. Incorporating interindividual variability is likely to improve clinical prognostics.

Commentary

The term “precision medicine” or “personalized medicine” has become ubiquitous in most areas of healthcare. Simply put, it is the “tailoring of medical treatment to the individual characteristics of each patient.” ¹ This approach involves the classification of patients into subpopulations based on their individual characteristics, including response to a specific treatment, to enable targeting of interventions to those most likely to benefit. This approach holds value at the individual level (e.g., a patient can weigh their chances of success with a therapy when considering its risk) and at a societal level (e.g., one can allocate resources rationally).

In the world of epilepsy, the term personalized medicine had largely evoked the increase in genetic markers of specific epilepsy etiologies, with the hope of translating this molecular knowledge to etiology-targeted therapies (e.g., gene therapies). ² More recently though, “personalized medicine” in epilepsy has expanded to other aspects of diagnosis and treatment, including predictors of drug and surgical treatment response. ³ Nomograms have emerged in the field of epilepsy surgery outcome prediction, ⁴ thus far drawing on basic clinical features readily extractable by clinicians. In their latest study, Lee et al. use innovative statistical methods to categorize 3T MRI multimodal features into accurate predictors of drug and surgery response and cognitive dysfunction. ⁵ There is a rationale for this approach—white matter microstructural anomalies have previously been correlated with specific cognitive phenotypes in temporal lobe epilepsy ⁶ and MRI morphometric analysis can predict surgical outcome, by using machine learning methods. ⁷

This study takes the concept of applying machine learning methods to imaging datasets to predict outcomes on an individual level one step further by investigating a number of MRI features previously shown to individually be relevant to temporal lobe epilepsy (TLE) phenotype and outcomes. ⁵ Eight-two patients with temporal lobe epilepsy underwent 3T MRI with structural and diffusion-weighted MRI sequences—of these, about 80% were drug-resistant. This proof-of-concept study employed rigorous methods to interpret well-defined neuroimaging characteristics that have previously been studied by the same group and validated by others, including neocortical and white matter morphological and microstructural anomalies. They employed an unsupervised machine learning method (Latent Dirichlet allocation) to develop a set of imaging characteristics either positively or negatively correlating with the outcome of interest—which in this study were surgical outcome (after selective amygdalo-hippocampectomy), drug response, and cognitive dysfunction. They then developed a list of characteristics variably co-expressed by individual patients: (1) ipsilateral hippocampal microstructural alterations and loss of myelin and atrophy; (2) bilateral paralimbic and hippocampal gliosis; (3) bilateral neocortical atrophy; and (4) bilateral white matter microstructural alterations. The categorization of an individual patient with each of these factors led to accurate prediction of drug response, surgical outcome, and cognitive dysfunction. Their findings are biologically plausible—markers of diffuse gliosis (involving hippocampi and neocortex) predicted poor surgical outcomes, which one might expect if the lesional zone extends beyond the resected area. The core innovation in this study lies in the movement past the correlation of a single marker with a single outcome, and demonstrates that machine learning has the ability to embrace the heterogeneity that is inherent to epilepsy and integrate this core feature of the disease toward reliable outcome prediction.

While the goal of the study was to embrace the heterogeneity of epilepsy, the cohort studied here was highly selected—only unilateral TLE was included and was defined using unequivocal unilateral scalp EEG onset or unilateral onset on stereo-EEG. Clean-cut unilateral temporal lobe epilepsy with hippocampal sclerosis has become increasingly rare. ⁸ Bilateral temporal epilepsy is common and likely a large driver of failure of resective surgery in this population—in fact, the predictors of poor outcome were bilateral abnormalities involving hippocampal and neocortical regions. In addition, no patients had a lesion other than hippocampal sclerosis, and a history of traumatic brain injury was exclusionary, but hippocampal sclerosis only accounts for about one-third of epilepsy surgeries. ⁹ Therefore, the generalizability to a broader set of temporal lobe epilepsies (let alone focal epilepsies) is limited. One would have hoped that the study design would have fully embraced the heterogeneity inherent to epilepsy as a disease by including a wider population of TLE. Future studies will likely employ similar methods, and include not only a less selected cohort of TLE, but also neocortical epilepsy. ¹⁰

Ultimately, we are also likely to use features beyond neuroimaging alone to personalize patient counseling, whether it involves likelihood of drug response, surgery success, or cognitive outcomes. The authors rightfully recognize that the investigation of a single MRI feature is not likely to change drastically our approach to outcome prediction and so expand to multimodal MRI features. But personalized outcome prediction will likely not be limited to a single testing modality—patients undergo evaluations spanning clinical domains, neurophysiologic testing, and neuropsychological investigations, all of which are considered in concert when counseling patients. Similarly, personalized medicine will need to mimic the clinical process, but improve upon it by removing human biases with unsupervised machine learning techniques. With this study, we are getting one step closer to that goal.

ORCID iD

Claude Steriade https://orcid.org/0000-0001-6799-9005