Abstract
Makhalova J, Medina Villalon S, Wang H, Giusiano B, Woodman M, Bénar C, Guye M, Jirsa V, Bartolomei F. Epilepsia. 2022;63(8):1942-1955. doi:10.1111/epi.17310 The virtual epileptic patient (VEP) is a large-scale brain modeling method based on virtual brain technology, using stereoelectroencephalography (SEEG), anatomical data (magnetic resonance imaging [MRI] and connectivity), and a computational neuronal model to provide computer simulations of a patient’s seizures. VEP has potential interest in the presurgical evaluation of drug-resistant epilepsy by identifying regions most likely to generate seizures. We aimed to assess the performance of the VEP approach in estimating the epileptogenic zone and in predicting surgical outcome. VEP modeling was retrospectively applied in a cohort of 53 patients with pharmacoresistant epilepsy and available SEEG, T1-weighted MRI, and diffusion-weighted MRI. Precision recall was used to compare the regions identified as epileptogenic by VEP (EZVEP) to the epileptogenic zone defined by clinical analysis incorporating the Epileptogenicity Index (EI) method (EZC). In 28 operated patients, we compared the VEP results and clinical analysis with surgical outcome. VEP showed a precision of 64% and a recall of 44% for EZVEP detection compared to EZC. There was a better concordance of VEP predictions with clinical results, with higher precision (77%) in seizure-free compared to non-seizure-free patients. Although the completeness of resection was significantly correlated with surgical outcome for both EZC and EZVEP, there was a significantly higher number of regions defined as epileptogenic exclusively by VEP that remained nonresected in non-seizure-free patients. VEP is the first computational model that estimates the extent and organization of the epileptogenic zone network. It is characterized by good precision in detecting epileptogenic regions as defined by a combination of visual analysis and EI. The potential impact of VEP on improving surgical prognosis remains to be exploited. Analysis of factors limiting the performance of the actual model is crucial for its further development.Objective:
Methods:
Results:
Significance:
Commentary
Some debates are ageless, like the chicken or the egg. Others can be measured in centuries, such as nature versus nurture. Many of these timeless debates will never be settled because both sides have merit. Today in surgical epileptology, the core debate is undeniable: Is it the focus or the network? In the mid-20th century, Jasper as well as Talairach and Bancaud were among the first to define the epileptogenic zone (EZ) as a focus where seizures originate. Lüders further described the EZ as the area of cortex that needed to be removed for complete seizure cessation. 1 Indeed, many drug-resistant epilepsy (DRE) patients achieve seizure freedom after confident localization and resection of EZs, supporting the idea of a focus. But with nearly half of patients continuing to have seizures after traditional diagnostic testing and surgical approaches, there is something we are missing with a focus-only perspective. Most of us have observed seemingly indistinguishable cases of well-localized DRE where surgery succeeds in one patient but fails in another. What is different? It has become increasingly clear that network connectivity is critical to understanding EZ pathophysiology. 2,3 A recent PubMed search for “epilepsy network connectivity” returned approximately 2000 citations, with about half published within the past 5 years. While the importance of network neuroscience in epilepsy is increasingly accepted, how can we apply network measures in the individual patient to improve surgical decisions? And can we use raw imaging and electrographic data we are already collecting, and examine them in a “new way” without throwing out the “old way?”
The highlighted study by Makhalova and colleagues is part of a growing body of work addressing these questions. 4 The authors examine the potential role of patient-specific modeling using the Virtual Epileptic Patient (VEP) in identifying EZ networks for surgery in DRE. 5 Virtual epileptic patient is a probabilistic modeling framework based on the Virtual Brain technology, described by the authors in a series of recent studies. The Bayesian approach uses personalized brain network models and machine learning methods to automatically sift through the clinical, imaging, and electrographic data and ultimately define EZs and propagation zones. In the present study, VEP modeling was retrospectively performed in 53 DRE patients incorporating stereo-EEG (SEEG) recordings, anatomic T1-weighted magnetic resonance imaging (MRI), and diffusion-weighed MRI (DWI). Virtual epileptic patient defined EZs were compared to clinically defined EZs. Patients were adults and older teens, about two-thirds of whom had a lesion on MRI, and just over one-half had temporal lobe or temporal-plus epilepsy. VEP models identified a total of 198 EZs across the 53 patients, 42% of which were not detected by clinical methods because the region (i) was not sampled by SEEG, (ii) was clinically designated as propagation zone but not EZ, or (iii) was deemed not involved on SEEG (about one-third in each category). Comparing VEP and clinical EZs, a mean 0.64 precision and 0.44 recall was noted, and improved performance was noted in nonlesional and temporal lobe cases. Just over one-half of patients had resection with postoperative outcomes available (50% Engel I), and the percentage of resected VEP EZs was significantly higher in the seizure-free group. Notably, the number of nonresected EZs detected by the model but not by clinicians was significantly larger in the patients with persistent postoperative seizures.
Overall, the study by Makhalova et al demonstrates a clear example of how patient-specific network models can be generated from multimodal clinical, imaging, and electrographic data that is already being collected clinically, and these models may aid in EZ localization. The authors also show detailed VEP model results in a few example patients, demonstrating how model results can be easily reviewed side-by-side clinical data to assist surgical decision-making. More favorable model performance in nonlesional patients is welcome, as these are individuals in whom novel localization tools are most critically needed. Also, in patients who continued to have postoperative seizures, the high percentage of VEP EZs that were not identified clinically suggests we may be missing true EZs in these individuals using traditional approaches alone. Prospective application of the model in the presurgical pipeline may lead us to take pause and investigate further prior to proceeding to surgery. Of note, the authors provide atlas parcellations online open source for evaluation in the community using external datasets.
Limitations of the study include a relatively small number of nonlesional and extratemporal patients in whom modeling is most needed, and a subset in whom surgical outcomes were not available for study. It is expected that these limitations may be overcome in the forthcoming multisite EPINOV trial evaluating VEP modeling approaches. 6 Concordance between VEP models and clinical EZs was relatively modest in certain patient subgroups, but as discussed here, this is in part because “gold standard” clinical interpretation is itself incorrect in some cases. Improved model performance (0.77 precision) in patients who achieved seizure freedom, in whom confidence in clinically designated EZs is intuitively higher, demonstrates this point. In one example patient, VEP modeling failed to identify a radiographically visible lesion as an EZ, but this simply demonstrates the value of side-by-side comprehensive review of model and clinical data. Of note, a potential limitation of machine learning techniques is our limited understanding of underlying pathophysiological features that are being detected by a seemingly “black box” approach. To improve interpretability, one can consider post hoc methods for data dimensionality reduction such as community detection or related techniques, as has been applied in recent studies. 7 Also, future models may consider incorporating functional MRI, which has been used to identify connectivity fingerprints that are related to seizure laterality 8 or postsurgical outcome. 9 Adding further clinical data points to the model may also be useful; one recent study demonstrated that a combination of electrographic and clinical data may improve performance in modeling surgical outcomes. 10
So, in surgical epilepsy, which is it: the focus or the network? Clearly, both concepts have merit, as focal epilepsy is a network disorder that can often be treated successfully by targeting critical nodes involved in seizure generation and propagation. A comprehensive approach using traditional diagnostic tests alongside patient-specific models that incorporate both brain network features and clinical data is likely to be most valuable moving forward. This field is growing rapidly, and thoughtful modeling studies such as the manuscript highlighted here are likely to help advance our field and ultimately improve clinical decisions and seizures outcomes in epilepsy surgery.
