Circadian Profile and Seizure Forecasting: Still Cloudy but With Chance for Sunshine

Commentary

The Circadian Profile of Epilepsy Improves Seizure Forecasting.

Karoly PJ, Ung H, Grayden DB, Kuhlmann L, Leyde K, Cook MJ, Freestone DR. 2017:140;2169–2182. doi:10.1093/brain/awx173.

It is now established that epilepsy is characterized by periodic dynamics that increase seizure likelihood at certain times of day, and which are highly patient-specific. However, these dynamics are not typically incorporated into seizure prediction algorithms due to the difficulty of estimating patient-specific rhythms from relatively short-term or unreliable data sources. This work outlines a novel framework to develop and assess seizure forecasts, and demonstrates that the predictive power of forecasting models is improved by circadian information. The analyses used long-term, continuous electrocorticography from nine subjects, recorded for an average of 320 days each. We used a large amount of out-of-sample data (a total of 900 days for algorithm training, and 2879 days for testing), enabling the most extensive post hoc investigation into seizure forecasting. We compared the results of an electrocorticography-based logistic regression model, a circadian probability, and a combined electrocorticography and circadian model. For all subjects, clinically relevant seizure prediction results were significant, and the addition of circadian information (combined model) maximized performance across a range of outcome measures. These results represent a proof-of-concept for implementing a circadian forecasting framework, and provide insight into new approaches for improving seizure prediction algorithms. The circadian framework adds very little computational complexity to existing prediction algorithms, and can be implemented using current-generation implant devices, or even non-invasively via surface electrodes using a wearable application. The ability to improve seizure prediction algorithms through straightforward, patient-specific modifications provides promise for increased quality of life and improved safety for patients with epilepsy.

“It's tough to make predictions, especially about the future,”

– Yogi Berra

As much as unpredictability of reward reinforces behavior, the unpredictability of punishment dramatically increases stress and anxiety (1). Such is the nature of seizures, where depression and anxiety are in large part due to the patients' sense of loss of control (2). Thus, making seizures predictable may have marked implications not only for the treatment of seizures but also for the improvement of the quality of life of patients.

Seizure prediction from EEGs has been attempted for the past 40 years with relatively modest success; no computer-assisted system has been widely adopted for clinical use. Reasons for this unexpected poor development are evident: there was a lack of long-term high-quality data, such as the data that can be obtained from intracranial recording; computational and storage powers were limited, and statistical and machine learning models were not well developed. The human brain simply is not as well wired to process time series data as it is for more visual imaging data.

Nonetheless, it is now an interesting era in which recent technological advances have allowed for some of these shortcomings to be addressed. With the advent of commercially available implantable intracranial recording devices, high-quality intracranial recording will likely become relatively widespread across a large number of patients with a variety of focal epilepsies (3). Machine learning, which itself had stagnated for long parts of the past half century (4), has more recently made remarkable progress, with some of the most advanced systems ironically mimicking the fundamental functions of human brain networks (5). As an example, the robustness of the most sophisticated algorithms has been demonstrated by the success of the recent kaggle.com competition sponsored by the National Institutes of Health and the American Epilepsy Society in which highly accurate algorithms for seizure detection in intracranial recordings were developed through a relatively inexpensive crowdsourcing infrastructure (6). It seems reasonable to expect clinically useful systems to be developed if a large amount of relevant high-quality data are provided. Thus, there is a greater need to understand what pertinent data would feed such systems. This article addresses one critical aspect of such data, namely, the circadian nature of human seizures.

In their article, the authors utilize the data from a landmark Australian study in which chronic intracranial recording devices were placed in 15 patients with epilepsy (7), of whom 9 were selected (including an additional patient not from the original cohort). They hypothesized that identifying individualized circadian patterns of seizures would substantially improve the forecasting model of seizures. Some of the theoretical and statistical concepts, such as metrics to determine the quality of forecasting, are derived from more mature weather forecasting statistics. Within this article, they elegantly elucidate many of the challenges of developing good forecasting models. Seizures, while devastating, are an extremely infrequent phenomenon, even in patients who are clinically deemed to have frequent seizures. Modeling such low-frequency events is particularly challenging as a forecast of low probability of seizures all the time would in fact be hard to beat. Developing a model that adds significant value to even such a cartoon scenario is a daunting task. Nonetheless, the authors do convincingly demonstrate that for about two-thirds of the patients, incorporating the time of seizure information, which is far less complex than the advanced mathematical models incorporated into the analysis of EEG signals themselves, substantially improves forecasting of seizure risk. The authors also correctly note that forecasting involves not only periods of high risk but also times of low risk, which may be just as important to the patient. The work depends on relatively sophisticated mathematical modeling, demonstrating that although it is conceptually easy to comprehend, incorporating circadian information into a useful model is a daunting task; other biologically relevant information may face similar challenges.

A few issues may dampen the enthusiasm behind these results. It has already been well established that seizures follow a circadian pattern in adults and children, which may be closely correlated to seizure onset zones (8, 9). Seizures in fact may be due to changes in more extensive brain networks (10). It may not be possible to include these aspects of seizures as one critical element of these chronic implantable devices has not yet been overcome—that the spatial resolution of implanted electrodes is extremely limited. In addition, one of their patients demonstrates the inherent issue of a circadian-based model. Patients and their physicians will react to obvious circadian patterns—it is common practice to increase antiepileptic doses into the evening regimen in patients who have predominantly nocturnal seizures—which alters the circadian pattern going forward. Nonetheless, there is a strong need for research into developing a comprehensive forecasting model that includes information regarding chronobiology, as well as other patient factors, such as medication schedule, age, concurrent medical issues, known anticonvulsant levels, and many other potential independent predictors of seizures; the authors themselves point out that even time from last seizure, a potentially critical element, had not been included. All this needs to be done, of course, without overfitting or overcomplicating the model.

Ultimately, the success of incorporating circadian patterns into forecasting models will depend on the answers to several questions: 1) Does such knowledge lead to an intervention that reduces seizures, rather than merely shifting them into more chronologically unpredictable events? 2) Is there improvement in antiepileptic drug–associated side effects by minimizing dosing during low-risk periods without turning them into high-risk periods? and 3) Does patient quality of life improve from receiving such forecasting notifications? My guess is that the answer to all of these questions is a resounding yes. While I suspect that clinically useful seizure forecasting will remain an extremely difficult task, there is reason to believe that substantial progress in the improvement of these models will be made. While forecasting seizures in patients with epilepsy is still somewhat cloudy, we can expect greater clarity in the future.