Introduction
From the early days of medicine, there has been efforts to predict the onset of a seizure (1). Sadly, these attempts at seizure forecasting, even up to the beginning of the 21st century, were no better than chance (2). However, a critical insight into the nature of epilepsy combined with the application of probability models and the development of wearable sensors now make seizure prediction a real possibility for all patients.
The importance of seizure prediction cannot be understated. Seizure prediction would radically improve the quality of life for the epileptic patient, allowing use of medication only when needed and/or providing sufficient preparation to avoid injury or death. Additionally, it would facilitate clinical trials to improve development of anti-seizure medication.
This blog will discuss the advancements to date that will in the future enable reliable seizure prediction for all.
Cyclic Nature of Seizures and Forecasting
The data from intracranial EEG recording devices highlighted the chronological pattern of seizures (3). Collected over a ten year period, seizure diaries combined with intracranial EEG activity (collecting data before, during, and after a seizure), characterized the repetitive patterns of seizures. Among a population of patients with epilepsy, these patterns occur less than 24 hours (ultradian), over a day (circadien), days and weeks (multidien) or yearly (circannual). Circadien and multidien patterns are the most common. Additionally, many smaller trials with continuous EEG recordings in patients with epilepsy confirmed the cyclic nature of seizures (1) (see also Blogs 8 and Blog 10).
The recognition of cyclic seizure eruptions provided impetus for forecasting. Specifically, the association between pre-ictal (before a seizure) activity and the seizure, itself, suggested potential predictability. Several clinical trials were completed to investigate this potential. The first trial on forecasting was the NeuroVista Trial (4). Using the Implanted Seizure Advisory System, it collected continuous intracranial EEG data in 15 patients for 2 years. For 9 patients without any adverse implantation effects, the Implanted Seizure Advisory System provided short term (minutes) seizure warnings at a performance level greater than chance (4).
A later clinical trial using a different intracranial monitor (RNS® System, NeuroPace, Inc, USA) implanted in 175 patients collected data over a longer period (5). In particular, this study developed a forecasting model in 18 patients and verified it in 157 patients over a 9 year period. Seizures could be predicted days in advance at a performance level greater than change in 66% of the 157 patients.
Application of Probability Models
Researchers considered that probability models were the most appropriate application to seizure forecasting. They considered the accuracy of the data, the quality of interpretation of the data and the time frame (days). This type of data did not allow a yes or no answer as might be expected on a clinical blood analysis. According to Baud et al.,(2023) these are basically the same reasons that weather forecasting converted to probability forecasting (6). Thus, the probability application rather than one that gives a definite outcome (deterministic) is best suited for EEG data. According to Baud et al., (2023),”by offering a graded assessment, probabilistic forecasting circumvents the core issues of specificity and sensitivity as well as false positives and negatives, which may raise stress for patients and create medico-legal issues, respectively.”
Non-Invasive Sensors
Intracranial implants require invasive surgery and are available for those with drug-resistant epilepsy. However, the preference by the majority of those with epilepsy would be some sort of wearable sensor with predictive capabilities. Therefore, the current focus is to develop wearable devices that detect cyclic biological phenomena that meaningfully relate to seizures in the same fashion that pre-ictal (prior to a seizure) epileptic activity does, as described above. As noted in the Figure below, development of such a sensor is a two-step process. First, the wearable sensor must accurately detect the seizure via a relevant bio-signal and secondly, the wearable sensor must provide a early and accurate warning signal.
One major working hypothesis is that electrical brain activity of seizure onset additionally affects specific functions of the autonomic nervous system (ANS) (7) . The ANS, in turn, controls activity of the cardiovascular system and temperature regulation. Thus far, heart rate, heart rate variability (time and electrical activity between beats) and thermoregulation such as electrical activity of the skin (e.g. sweat gland activity) and also the sleep-wake cycle, are targeted candidates. Interestingly, these bio-signals exhibit circadian and multidien cyclical relationships as do seizures.
Feasibility
The feasibility of this hypothesis has been tested. In one interesting study (8), data was collected over several days from 9 epileptic patients implanted with scalp electrodes (less invasive than intracranial implants), standard EEG leads and a chronic EKG recorder. Machine learning generated algorithms to forecast a seizure. Scalp implanted electrodes were the most reliable as seizure predictors. However, EKG recordings that assessed more than just heart rate proved prediction as good as a 21 channel EEG. Thus, seizure prediction seemed possible with cardiovascular bio-signals.
Clinical Trials
Large clinical trials have evaluated two wearable sensors that are part of step one, that is, detection of a seizure with ANS bio-signal (see Figure). Those evaluated are the armband Nightwatch system and the Empatica wristwatch device. The Nightwatch system is an upper arm bracelet that measures accelerometry (detecting physical activity) and all aspects of heart changes in patients with nighttime seizures (9,10). Its sensitivity ranges from 86% in adults and 75% in pediatric patients. The Empatica wristwatch device exhibited a higher sensitivity in adults and children but has a higher false alarm rate than the Nightwatch system (11). Seizure detection was less than a minute. Many other devices are in early clinical testing (7).
Studies of wearable sensors with forecasting capabilities are just beginning. Wearable sensors must identify the best bio-signal that aligns with pre-ictal changes to make an accurate prediction. Study results are positive but trials are in early stages with small number of volunteers (7). Many studies focus on heart rate cycles as relevant predictors. However, in a study of 69 patients, seizure predictions, better than chance, occurred in approximately half of the patients (with 30 minute lead time). In this case, multiple bio-signals (heart rate, temperature, movement and electrodermal activity) were detected with a multimodal wristwatch (8).
Conclusions
The recognition of the cyclic nature of seizure activity provides the potential for seizure prediction. Better than chance seizure prediction is already a reality with intracranial implant systems. The next step is seizure prediction with a noninvasive devices e.g. a wearable sensor that use bio-signals relating to seizure onset. To date, wearable sensors can detect seizures better than chance. In the future, wearable sensors will also predict seizures better than chance.
From: Miron G, Halimeh M, Jeppesen J, Lodenkemper T, Meisel C. Autonomic biosignals, seizure detection, and forecasting. Epilepsia. 2025 Sep;66 Suppl 3(Suppl 3):25-38 doi: 10.1111/epi.18034.
References
1. Baud MO et al. In: Noebels JL, Avoli M, Rogawski MA, Vezzani A, Delgado-Escueta AV, editors. Jasper’s Basic Mechanisms of the Epilepsies. 5th edition. New York: Oxford University Press; 2024. Chapter 15.
2. Baud MO et al. Seizure forecasting: Bifurcations in the long and winding road. Epilepsia. 2023 Dec;64 Suppl 4(Suppl 4):S78-S98. doi: 10.1111/epi.17311
3. Leguia MG et al. Seizure Cycles in Focal Epilepsy. JAMA Neurol 78, 454–463, doi: 10.1001/jamaneurol.2020.5370 (2021).
4. Cook MJ et al. Prediction of seizure likelihood with a long-term, implanted seizure advisory system in patients with drug-resistant epilepsy: A first-in-man study. Lancet Neurol. 2013; 12(6):563–71.
5. Proix T. et al. Forecasting seizure risk in adults with focal epilepsy: a development and validation study. Lancet Neurol 20, 127–135, doi: 10.1016/S1474-4422(20)30396-3 (2021).
6. Lorenz E. Deterministic Nonperiodic Flow. J Atmos Sci. 1963; :130–41.
7. Miron G et al. Autonomic biosignals, seizure detection, and forecasting. Epilepsia. 2025 Sep;66 Suppl 3(Suppl 3):25-38 doi: 10.1111/epi.18034.
References (additional)
8. Meisel C, Bailey KA. Identifying signal-dependent information about the preictal state: a comparison across ECoG, EEG and EKG using deep learning. EBioMedicine. 2019;45:422–31.
9. Arends J et al. Multimodal nocturnal seizure detection in a residential care setting: a long-term prospective trial. Neurology. 2018;91(21):e2010–e2019.
10. Van Westrhenen A, et al. Dutch TeleEpilepsy Consortium. Multimodal nocturnal seizure detection in children with epilepsy: a prospective, multicenter, long-term, in-home trial. Epilepsia. 2023;64(8):2137–52.
11. Onorati F et al. Prospective study of a multimodal convulsive seizure detection wearable system on pediatric and adult patients in the epilepsy monitoring unit. Front Neurol. 2021;12:724904.
12. Meisel C et al. Machine learning from wristband sensor data for wearable, noninvasive seizure forecasting. Epilepsia. 2020;61(12):2653–66.