Abstract
Presurgical Thalamic “Hubness” Predicts Surgical Outcome in Temporal Lobe Epilepsy
He X, Doucet GE, Pustina D, Sperling MR, Sharan AD, Tracy JI. Neurology 2017;88:2285–2293.
OBJECTIVE: To characterize the presurgical brain functional architecture presented in patients with temporal lobe epilepsy (TLE) using graph theoretical measures of resting-state fMRI data and to test its association with surgical outcome. METHODS: Fifty-six unilateral patients with TLE, who subsequently underwent anterior temporal lobectomy and were classified as obtaining a seizure-free (Engel class I, n = 35) vs not seizure-free (Engel classes II–IV, n = 21) outcome at 1 year after surgery, and 28 matched healthy controls were enrolled. On the basis of their presurgical resting-state functional connectivity, network properties, including nodal hubness (importance of a node to the network; degree, betweenness, and eigenvector centralities) and integration (global efficiency), were estimated and compared across our experimental groups. Cross-validations with support vector machine (SVM) were used to examine whether selective nodal hubness exceeded standard clinical characteristics in outcome prediction. RESULTS: Compared to the seizure-free patients and healthy controls, the not seizure-free patients displayed a specific increase in nodal hubness (degree and eigenvector centralities) involving both the ipsilateral and contralateral thalami, contributed by an increase in the number of connections to regions distributed mostly in the contralateral hemisphere. Simulating removal of thalamus reduced network integration more dramatically in not seizure-free patients. Lastly, SVM models built on these thalamic hubness measures produced 76% prediction accuracy, while models built with standard clinical variables yielded only 58% accuracy (both were cross-validated). CONCLUSIONS: A thalamic network associated with seizure recurrence may already be established presurgically. Thalamic hubness can serve as a potential biomarker of surgical outcome, outperforming the clinical characteristics commonly used in epilepsy surgery centers.
Commentary
The study by He et al., showed that thalamic hubness may serve as a potential biomarker for seizure outcome after anterior temporal lobectomy. But, before the results of the present study are considered, we need to understand two overarching concepts—the current thinking about the role of thalamus in focal seizure generation and maintenance, and the basics of graph theory.
The mechanistic role of thalamus in generation and maintenance of focal and generalized epilepsies was initially proposed in the 1950's by Penfield based on his work and the work of others, including Jasper et al. (1). While initially considered important for seizure generation and maintenance in generalized epilepsies, thalamic participation as a hub in the temporal lobe epilepsy (TLE) network has been increasingly recognized (2). This and other imaging studies have confirmed the thalamic involvement previously observed in animal and human electrophysiologic recordings in the maintenance of consciousness during seizures and the relationship between surgical outcomes and the presence or absence of thalamic involvement (3). Further, a recent structural connectivity study indicated that thalamic bilateral dorsomedial and pulvinar region atrophy and alternations in thalamotemporal connectivity as measured with diffusion tensor imaging (DTI) predicted surgery failure (4). Thus, thalamic participation in seizure generation and maintenance based on the available literature is unquestionable. But, how thalamic involvement relates to seizure intractability and how functional connectivity fits into this picture is unclear, and this is where the paper by He et al. brings some needed clarity.
The second issue is to understand the concept of ‘graph theory’—a concept that is likely esoteric to a practicing neurologist. In general, it is clear that complex systems, like those of a human brain, consist of elements that interact with each other. These systems exhibit dynamic behaviors that cannot be deduced from simple analysis of isolated regions. This is where in some situations graph theory may come into play as it allows for mathematical analysis and description of complex systems (5,6). Concepts used in graph analyses include ‘hubs’ and ‘hubness’, ‘edges’, ‘betweenness’, ‘efficiency’, ‘centrality’, ‘resilience’, and so on. These complex concepts may be difficult to understand by a nonmathematician. Thus, let's try a simple travel analogy to explain graph theory and the terms used by He et al. For example, let's think through five different airports and the relationships between them. Considering efficiency of the network, flying directly from Birmingham, AL to Washington, D.C. would be considered efficient while taking three additional layovers in San Francisco, CA, Atlanta, GA, and Chicago, IL would be considered inefficient. All five of the mentioned airports are hubs in the network but the hubness of Birmingham, AL with fewer connections is lower than the hubness of San Francisco, CA or Atlanta, GA that have many more connections. Thus, the importance (or hubness) for the network of, for example, the Atlanta, GA airport is higher than that of the Birmingham, AL airport. While these five airports form a network, they have different centralities—a concept that indicates the importance for the cohesiveness of this network or in this case, the “richness of connections.” Betweenness is another concept—a measure of centrality based on the shortest path between this and other nodes—using our travel analogy, the airport with the best betweenness property is the airport that has the highest number of shortest connections between other airports and highest throughput—or, in the network examined here, the node that has the most control over the network. Next, the “eigenvector centrality”—a measure of the importance one node has in relationship to the other nodes of the network—or, using the air travel analogy, it is the importance of having or not having this particular airport as part of the entire network. Finally, the authors examined “global efficiency” (or integration) of the network—in our travel analogy the question this measure addresses is: the shutdown of which of the above airports would result in most disruption?
So, back to the study—there are several strengths of the graph theoretical approach used by He et al. The authors have shown that thalamic network associated with seizure generation may be established in some patients (greater number and more important connections) before epilepsy surgery and if so, the presence of these connections and thalamic participation in the network may be a potential biomarker for predicting surgical outcome in patients with temporal lobe epilepsy. If the thalamus is involved, then removing the more peripheral node—medial temporal structures—may not result in seizure freedom. Thus, using graph theory terms, not seizure-free patients had increased hubness in ipsi- and contralateral thalami (degree and eigenvector centralities). Simulated removal of the thalamic hub (not that anyone is suggesting thalamic resection) reduced network integration in not seizure-free patients, providing additional theoretical evidence in support of thalamic involvement in seizure network integration and in seizure generation and maintenance.
This study has many common features with several recent studies that used sophisticated neuroimaging techniques to predict anterior temporal lobectomy outcomes (7). All in all, the recent studies have relatively low-to-moderate numbers of subjects enrolled at one site with data processed using techniques typically not accessible to a clinician—techniques that use sophisticated mathematical algorithms or data collection methods that make these techniques useful research tools but make it very difficult to implement them in “real-life” situations. There are other limitations of this study. For one, the thalamus is a large structure so using the whole thalamus for the analyses may be an oversimplification of the importance of the entire structure for seizure generation and maintenance. In fact, in patients with generalized epilepsies, different thalamic nuclei have been shown to have different time course of the blood oxygenation level dependent (BOLD) signal changes related to generalized spike and wave discharges (8) and corticothalamic connectivity has been shown to be different depending on the location of the thalamic BOLD signal (9). Further, a recent trial of deep brain stimulation selected anterior thalamic nucleus rather than “thalamic electrode placement” questioning the specificity of using the “thalamic hubness” for outcome prediction rather than hubness of a specific thalamic nucleus (10). Another criticism is that this study focused on a group of patients that has the highest rate of seizure freedom after surgery—temporal lobe epilepsy. It would be more interesting and telling if the analyses proposed by He et al. could be applied to and results replicated in patients with extratemporal neocortical epilepsies.
We need to take a step back and review what is available in the neuroimaging literature. Most imaging studies are small to moderate in size, use techniques unique to one lab, address specific groups of patients or specific surgical techniques, lack replication, and so on. Maybe it is time to create large database of epilepsy patients with various pathologies or national registries of epilepsy surgery patients akin to FITBIR (https://fitbir.nih.gov/) for traumatic brain injury to evaluate the contributions of these sophisticated techniques to the overall outcomes of the patients? After all, is another small study going to address the need of a specific patient who is somewhat different from the patients enrolled in that particular effort?