Abstract
Khambhati AN, Chang EF, Baud MO, Rao VR. Nat Med. 2024. doi: 10.1038/s41591-024-03149-6. Online ahead of print. Seizures in people with epilepsy were long thought to occur at random, but recent methods for seizure forecasting enable estimation of the likelihood of seizure occurrence over short horizons. These methods rely on days-long cyclical patterns of brain electrical activity and other physiological variables that determine seizure likelihood and that require measurement through long-term, multimodal recordings. In this retrospective cohort study of 15 adults with bitemporal epilepsy who had a device that provides chronic intracranial recordings, functional connectivity of hippocampal networks fluctuated in multiday cycles with patterns that mirrored cycles of seizure likelihood. A functional connectivity biomarker of seizure likelihood derived from 90 s recordings of background hippocampal activity generalized across individuals and forecasted 24 h seizure likelihood as accurately as cycle-based models requiring months-long baseline recordings. Larger, prospective studies are needed to validate this approach, but our results have the potential to make reliable seizure forecasts accessible to more people with epilepsy.
Commentary
Are my medications working? Is it safe for me to go out today? When will my next seizure strike? These are some of the agonizing questions that people with drug-resistant epilepsy (DRE) often ask themselves. The unpredictability of seizures in DRE is one of the most challenging aspects of this disabling disorder. Patient survey data suggest that “anticipation of the next seizure” is the single greatest cause of anxiety in DRE, fueling comorbid mood disorders in this population. 1 Thus, seizure prediction is somewhat of a “holy grail” in human epilepsy research, even sparking competition within anti-epilepsy organizations. Yet, progress in the field of seizure forecasting remains elusive, likely in part due to complex large-scale network underpinnings that underlie peri-ictal dynamics in the epileptic brain. 2 Much of the work in seizure prediction utilizes scalp or intracranial electroencephalography (EEG) data, seeking to uncover signature electrographic changes that foretell an imminent ictal event. 3 While intracranial recordings offer the most direct measure of neural activity, data collected in the epilepsy monitoring unit are short-lived and from an environment that does not resemble a patient's home conditions. However, with the responsive neuromodulation (RNS) system, we now have access to long-term electrocorticography (ECoG) collected in an individual's natural environment via a chronically implanted neuromodulation device. Might these daily intracranial recordings offer a unique window to develop models that accurately predict seizure probability?
In the presently highlighted study, Khambhati et al 4 aimed to leverage hippocampal network activity to forecast epileptic seizures in 15 bilateral mesial temporal lobe epilepsy (mTLE) patients implanted with RNS. The project focused on identifying functional connectivity (FC) patterns within and between the hippocampi during brief 90 s interictal periods. By correlating these FC patterns with multidien cycles of interictal epileptiform activity (IEA) extracted using a Morelet wavelet transform, the authors sought to develop a biomarker capable of predicting seizure likelihood over a 24 h horizon. Investigators hypothesized that phase-related FC biomarkers conserved across individuals could be leveraged to forecast seizure likelihood in out-of-sample patients using a generalized linear model. This FC-IEA phase template map demonstrated increased wide-range delta–theta band FC during the rising phase, elevated long-range gamma connectivity during the peak, reduced wide-range broadband FC during the falling phase, and increased local broadband and long-range delta–theta connectivity at the trough. In the end, the authors found that prediction of seizure occurrence with an FC-based model was on par or better than benchmark cycle-based models that require months of baseline data and that forecasting of seizures was significantly better and potentially clinically useful with an IEA-based model.
Overall, this is an elegantly designed and encouraging study showing that a brief 90 s clip of hippocampal ECoG from only 4 bipolar pair channels yields seizure forecasting as good as noncausal IEA cycle modeling. The work builds upon prior research by this group in a larger and more diverse RNS patient cohort, which demonstrated that seizure risk varies through multidien rhythms over several years of treatment. 5 Specific analytical strengths of the present study include rigorous permutation testing and balancing of clips used for model generation. Importantly, the work further demonstrates that seizure propensity is not a haphazard phenomenon free of discernible patterns, but that epilepsy can be a somewhat predictable cyclical disorder—a finding consistent with “seizure clusters” often reported by patients.
Limitations of the presently highlighted investigation include a small sample size and its restriction to bilateral mTLE patients. Thus, validation in larger and more heterogeneous cohorts of DRE patients treated with RNS remains a future endeavor. Also, correlations between FC and the IEA phase in the authors’ model were statistically significant but relatively modest. Reliable prediction depended heavily on the phase of the IEA cycle, which had a moderate area under the curve of 0.72 ± 0.03, and models to predict seizure risk directly from FC will be a lofty but worthwhile future goal. Despite these limitations, progress in the field of seizure forecasting propelled by the study is palpable and valuable.
Paradigms to predict seizures on a circadian basis, such as that reported by Khambhati et al, 4 would help reduce patient fear and anxiety in DRE and may guide daily adjustments in behavioral, neuromodulatory, and pharmaceutical treatment strategies. However, multidien electrographic seizure forecasting in DRE does not portend accurate real-time detection of pre-ictal network states that will be necessary for the next generation of closed-loop neuromodulation approaches. The success of current RNS treatment paradigms is favorable but also imperfect. While the premise of RNS technology involves real-time detection and termination of ictal rhythms, recent work suggests that termination of individual seizures is often unsuccessful, and long-term neuromodulatory impacts contribute substantially to the clinical benefit of RNS. 6 The ideal neuromodulation strategy in DRE would prevent all seizures, and will likely require adaptive, network-based prediction and prevention of pre-ictal states.
Moving toward a novel adaptive neuromodulation system for DRE will require addressing several questions: where to stimulate, when to stimulate, and how to stimulate the epilepsy network. Recent intracranial EEG studies suggest that network-based analyses may help localize sites for treatment targeting, and at rest, these epileptogenic hubs often demonstrate strong inward connectivity from more normal brain nodes, contrasted by lower outward connectivity.7,8 It is possible that this connectivity signature in epileptogenic nodes reflects an inhibitory network phenomenon that helps counteract seizure states but connectivity patterns may then shift during the peri-ictal period. There is preliminary evidence that these shifts from baseline interictal to a pre-ictal network state can be detected and tracked in individual patients using intracranial electrophysiology, and that network shifts may be induced with lower frequency and lower amplitude pulse stimulations than are typical in current neuromodulation strategies. 9
Finally, whereas the highlighted study by Khambhati et al 4 used a partially hypothesis-driven method to identify biomarkers for seizure prediction, the field of data-driven machine learning for seizure forecasting is rapidly expanding. 10 While hypothesis-driven approaches may be more likely to scale across patients with a common type of DRE, such as bilateral mTLE, it is possible that artificial intelligence methods will perform best when trained on large, heterogeneous electrographic datasets. Limitations of these machine learning methods include data overfitting, under-sampling, and computationally intense processes that would be challenging to achieve with a resource-limited and energy-limited implanted neuromodulation device. The future of adaptive neuromodulation may benefit from a mixed approach that balances scalability and personalization.
Seizure forecasting remains elusive and challenging, but is nonetheless critical in epilepsy. It has significant implications for patient warning, treatment guidance, and development of network-based adaptive neuromodulation therapies. Nevertheless, progress in the field appears to be substantial in recent years, hopefully moving us one step closer to that “holy grail.”
