Feasibility study validates use of epileptiform activity for predicting seizures

Data from implanted devices about interictal epileptiform activity over the course of days enabled researchers to forecast seizure activity in advance, according to the results of a feasibility study published in The Lancet Neurology.

“People with epilepsy live in a state of uncertainty and the constant threat of seizure has implications for personal safety, independence and psychological well-being,” the researchers wrote. “Reliable methods to anticipate seizures would mark a paradigm shift in clinical epilepsy, mitigating uncertainty and enabling time-varying, risk-based seizure prevention strategies.”

However, despite progress in the field of seizure prediction, “such methods remain elusive,” Timothée Proix, PhD, of the faculty of medicine in the department of basic neurosciences at the University of Geneva in Switzerland and the department of neuroscience at the Carney Institute for Brain Science at Brown University in Rhode Island, and colleagues continued. They hypothesized that seizure probability could be determined by alignment of cyclical influences at multiple timepoints and by the temporal distribution of prior seizures.

Proix and colleagues conducted a feasibility study to assess models that include elements of cyclical influences and temporal distribution to produce seizure risk forecasts. The feasibility study served as a precursor for larger prospective clinical trials meant to validate these forecasts, according to the researchers.

The feasibility study involved a retrospective analysis of chronic EEG (cEEG) data recorded on implanted devices in adults with drug-resistant focal epilepsy. Patients were required to have had 20 or more electrographic seizures (development cohort) or self-reported seizures (validation cohort). In all patients, the device recorded interictal epileptiform activity (IEA; 6 months of continuous hourly data), and the researchers used the fluctuations to help estimate varying seizure risk. Further, point process statistical models trained on initial portions of each patient’s cEEG data in both cohorts provided forecasts of seizure probability. Proix and colleagues tested those forecasts on subsequent unseen seizure data and evaluated that against surrogate time series.

The percentage of patients whose forecasts of seizure probability demonstrated improvement over chance (IoC) served as the primary outcome.

Proix and colleagues screened 72 patients for the development cohort and 256 patients for the validation cohort. The study ultimately included 18 patients in the development cohort and 157 patients in the validation cohort.

Baseline characteristics were similar in both cohorts. Median age was 38 years (interquartile range, 32-51 years) in the development cohort and 35 years (interquartile range, 25-43 years) in the validation cohort. The development cohort included eight women (44%) and the validation cohort included 74 (47%). Mesiotemporal epilepsies were predominant, according to the study findings.

Models that included information about IEA cycles over days (ie, multidien) alone resulted in daily seizure forecasts for the next calendar day, with IoC in 15 patients in the development cohort (83%) and 103 patients in the validation cohort (66%). The forecasting horizon could be expanded up to 3 days while retaining IoC in two of 18 patients (11%) in the development cohort and 61 of 157 patients (39%) in the validation cohort. Forecasts with a shorter horizon of 1 hour, which were possible only for electrographic seizures in the development cohort, demonstrated an IoC in all 18 patients.

Proix and colleagues also examined forecast performance for patients with IoC using two complementary metrics: area under the curve (sensitivity vs. corrected time in warning), to assess the value of a forecast given the time spent in warning, and Brier skill score, to examine how well a forecast performed relative to a naive reference strategy (BSS = 1 for a perfect forecast and BSS = 0 for no improvement over a random predictor).

For daily forecasts with models that included information about multidien phases, median AUC was 0.74 (IQR, 0.69-0.8) for electrographic seizures (development cohort) and 0.7 (0.65-0.75) for self-reported seizures (validation cohort). Median BSS was 0.23 (IQR, 0.18-0.31) and 0.13 (IQR, 0.05-0.2), respectively.

The study provided forecasts for electrographic seizures and self-reported seizures, the latter of which represent the “gold standard” of assessment in epilepsy practice and clinical trials, up to three days in advance, Proix and colleagues wrote. That horizon for personal seizure risk stratification, to the researchers’ knowledge, “is unprecedented.”

“... Our results corroborate an emerging view that seizures are not entirely random events,” the researchers wrote. “Given the large sample size, the present results validate and extend our previous findings based solely on electrographic seizures, and they suggest the generalizability of multiscale cyclical biomarkers in epileptic brain activity to forecast clinically relevant seizures over long horizons. Furthermore, our study indicates that seizure forecasting could be feasible in the treatment of refractory focal epilepsy with an FDA-approved medical device in widespread clinical use in the U.S.”

Forthcoming prospective clinical trials are planned that will evaluate the benefits of “replacing a state of constant subjective uncertainty about upcoming seizures with a continuum of quantified uncertainty (a forecasted probability) at different horizons,” according to the researchers.

In a related editorial, Mark J. Cook, MD, director of the Graeme Clark Institute, Sir John Eccles Chair of Medicine and director of clinical neurosciences at St. Vincent’s Hospital in Melbourne, wrote that the study by Proix and colleagues “confirms the initial observations” regarding long and short cycles of seizure activity and their value in predicting new-onset seizures.

“The authors observe the importance of EEG over self-reported episodes in seizure prediction to overcome the little discussed problem of overreporting of seizures in patients during in-hospital monitoring. This problem might become more widely appreciated though ambulatory video-EEG systems,” Cook wrote. “The authors also highlight that seizure occurrence is best conceptualized as a continuously fluctuating background risk best suited to a probabilistic forecasting framework, rather than a deterministic one.”

The idea of forecast framing has implications for the development of prediction systems and the application of these strategies in clinical practice, as the goal for many patients with seizures “is to receive a more definitive notification from a seizure prediction system than is feasible.”

While “many unknowns remain,” according to Cook, including the underlying drivers of seizure patterns and whether or not the prediction of seizures will have a meaningful impact for patients, these factors — once addressed — “could be developed into systems that lead to improved management of epilepsy, with seizure detection systems linked to forecasting of events and safety management processes,” he wrote.