Out of one, how many? Subtyping in epilepsy

Abstract

This scientific commentary refers to ‘Identification of different MRI atrophy progression trajectories in epilepsy by subtype and stage inference’ by Xiao et al. (https://doi.org/10.1093/brain/awad284).

This scientific commentary refers to ‘Identification of different MRI atrophy progression trajectories in epilepsy by subtype and stage inference’ by Xiao et al. (https://doi.org/10.1093/brain/awad284).

Subtyping in epilepsy has long been fruitful, providing a common language among the epilepsy community.¹ Generalized versus focal epilepsies, and subdivisions based on focal seizure onsets and lesions (e.g. mesial temporal sclerosis) influence choices between medical management, destructive surgical procedures and more recently neuromodulation. An emerging dimension of this decision-making is whether ongoing epilepsy is associated with progressive pathology, which can be loosely termed neurodegenerative.^2,3 A destructive surgical treatment may reduce neurodegenerative effects associated with prolonged epilepsy³ and ultimately reduce mortality⁴ but also presents risks (e.g. significant cognitive decline). Indeed, surgical outcomes are variable, with 10-year seizure freedom following even long-used surgical procedures such as anterior temporal resection as low as 49%.⁵ Therefore, evolving and refining our understanding of how subtyping may improve the targeting of individual patients among the increasing treatment options would be a boon. In this issue of Brain, Xiao and co-workers⁶ take a step along this path by tackling questions of both subtyping and progression in a large and heterogenous cohort of patients with epilepsy using an emerging machine-learning technique.

The technique is named subtype and stage inference (SuStaIn) and it aims to uncover data-driven disease phenotypes with distinct temporal progression patterns from cross-sectional patient imaging. The authors obtained a remarkable database of T₁-weighted imaging, including 1068 scans (879 patients with epilepsy and 189 healthy controls) across two centres. They structured their input to SuStaIn in two important ways. First, though they separated patients into focal and generalized epilepsy, they combined data within these groups for input into SuStaIn, implicitly eschewing existing classifications in the search for common patterns of pathology progression (Fig. 1). Second, they separated the two centres into a discovery cohort (696 patients) and a validation cohort (183 patients), to probe the replicability of the identified patterns of progression.

Figure 1.

Approaches to subtyping in epilepsy. The top half of the figure (Clinical Classifications) illustrates the patients included by Xiao et al.,⁶ divided into focal epilepsy and idiopathic generalized epilepsy (IGE) subtypes. These subtypes are well established in the epilepsy field, with judgements made by clinical team consensus. The bottom half of the figure represents the reclassification of these patients using the subtype and stage inference (SuStaIn) approach using T₁-weighted structural imaging from each patient. Focal and IGE patients were pooled into SuStaIn separately, with three resulting subtypes identified. Two were common across focal epilepsy and IGE (cortex subtype and basal ganglia subtype) while one was unique to focal epilepsy (hippocampus subtype).

In the discovery cohort, the authors identified three subtypes of epilepsy disease progression, beginning in the cortex (46% of patients), basal ganglia (30%) or hippocampus (24%). The cortical pattern, present in both the focal and generalized groups, affected firstly frontal-temporal regions, before spreading to parietal-occipital regions and affecting subcortical regions only at the latest stages. The basal ganglia subtype, again present in both the focal and generalized groups, began in the globus pallidus, before spreading throughout the basal ganglia and then to the cortex in the latest stages. The hippocampal subtype, uniquely present only in the focal group, started in the hippocampus, spreading to the basal ganglia and finally onto the cortex. Importantly, there was no evidence of merging events between subtypes. The authors argue that these subgroupings, mainly consistent across both focal and generalized epilepsies, should prompt a reconceptualization of epilepsy as ‘dynamic disorders associated with a deterioration of brain structure’.

The authors next investigated how their novel subtypes relate to clinical characteristics and treatment outcomes. Finding associations was more successful in the focal group than the generalized group. In the focal group, across subtypes the duration of epilepsy correlated with the predicted stage, giving credence to the idea of a degenerative aspect, but age of onset was not correlated with staging. The authors also identified disease burden principal components, finding that the basal ganglia subtype was associated with greater disease burden. Further, of the focal patients that had epilepsy surgery, patients with the basal ganglia subtype were less likely to have an Engel score I–II (50%), whereas the cortical (81.3%) and hippocampal (73.7%) subtypes had better outcomes. Taken together, this suggests that the basal ganglia subtype in focal epilepsy is associated with poorer prognosis for standard treatments.

In addition to clinical correlates, it is also important to show generalizability. Here, the authors are able to leverage not just the generalizability of patterns of progression across focal and generalized epilepsy mentioned above, but also across epilepsy centres. When comparing discovery and validation cohorts, they found broad generalizability but also some subtle differences. Importantly, the validation cohort replicated the three subtypes and the proportions. In the focal validation group, the authors replicated the relationship of stage to duration as well as the observation of worse surgical outcomes in the basal ganglia group. Interestingly, the most prominent difference between datasets was in the pattern of spread in the cortex group. In the validation group, frontal pathology preceded temporal pathology. The authors hypothesized that this was due to differing proportions of temporal and frontal lobe epilepsy in the discovery and validation cohorts. Returning to the discovery group, they split the focal group into temporal and extra-temporal seizure onset, with the temporal-onset cortical group demonstrating temporal pathology onset before frontal pathology and the opposite pattern in the extra-temporal group. This suggests that seizure onset region still maintains its place as an initial site of pathology even within the context of broader degenerative patterns.

Outside evidence for generalizability comes from a paper based on a temporal lobe epilepsy-only cohort, which used event-based modelling, a technique that orders disease progression but does not attempt to subtype. In this temporal lobe-only cohort, the authors found a pattern of pathological progression nearly mirroring the hippocampal group from Xiao et al.⁶ Pathology began in the hippocampus, spread to the thalamus, then diffusely onto widespread cortical regions.⁷

Parsing the results of Xiao et al.⁶ shows one immediately provocative and testable hypothesis. Patients judged by MRI to present with mesial temporal sclerosis (MTS) were spread throughout the three groups (cortical: 32%, basal ganglia: 10%, hippocampal: 58%). This suggests that ∼42% of MRI-identified cases of MTS are secondary to preceding non-hippocampal pathology. Given the important role of MTS in clinical decision-making (e.g. laser ablation versus resection), if MTS proves often to be secondary, that would be an important clinical consideration. The authors highlight the fact that 10-year seizure freedom can be as low as 49% following anterior temporal resection,⁵ which may suggest that the removal of the hippocampus does not remove a broader epileptogenic network in the non-seizure free patients even in MTS cases, thus resulting in poor seizure control. Further studies are necessary to understand whether this data-generated hypothesis has a biological, and clinical, basis.

Stepping back from this detailed work focused on T₁-weighted neuroimaging and SuStaIn, complementary lines of research also seek to evolve and refine subtypes, using additional neuroimaging modalities, behavioural phenotypes and genetics.⁸ This naturally leads to questions about which models predominate, and why? Several perspectives are at odds with the results of Xiao et al.,⁶ which may profitably be explored in future research. First, in studies in which phenotype is based on cognitive performance, there is always a group that is nearly indistinguishable from controls, with no group differences in cognitive performance and a lack of brain pathology.^8,9 The SuStaIn model allows for ‘Stage 0’ patients in which there are no abnormalities; however, in this study not a single patient was assigned to this category.

Second, the authors’ previous work^2,3 and their current results place considerable emphasis on epilepsy duration, with epilepsy suggested to be neurodegenerative. These studies were largely focused on cortical measures. Cortical thickness begins to decrease from about 1 year of age, presenting a linearly declining signal. In contrast, white matter continues to develop through adolescence, leading to a potentially non-linear role of age of onset and disrupted development in white matter as opposed to linearly accelerated decline.¹⁰ A potential way to begin integrating across methodologies and domains would be an ‘open competition’ in which all studies use the same dataset, such as ENIGMA-Epilepsy. This would allow comparisons of similarity and perhaps more interestingly, differences in subtyping.

Giving full meaning to these intriguing discussions will depend on how they jump the gap from the laboratory to the clinical environment. For instance, how do the proposed classifications meld with the International League Against Epilepsy’s taxonomic system for defining seizure types, syndromes and epilepsy aetiologies? In a decade or two, will case conferences be discussing cortical versus basal ganglia subtypes alongside, or perhaps instead of, frontal versus temporal lobe epilepsy? Like the translation of concepts such as ‘epilepsy as a network disease’ into clinical thinking, this will take much replication and an organic dialogue between the clinical and research teams.

Funding

E.K. was supported by NINDS (K01NS12483). A.R. was supported by a NINDS Diversity Supplement (R01 NS120976), Clinical Research Grant from the National Academy of Neuropsychology, and a Postdoctoral Diversity Enrichment Program from the Burroughs Wellcome Fund.

Competing interests

The authors report no competing interests.