The ability to track seizures has a number of potential benefits: It could allow physicians to better determine optimal dosing and timing of medication, as well as enable timely interventions to help prevent impending seizures.
Traditionally, electroencephalography (EEG) and electrocorticography have been used to evaluate and forecast seizures.
One such device, a biosensor wristband developed by Empatica, records autonomic nervous system features, patient movement, and other vital signs and parameters.
To learn more, Tobias Loddenkemper, MD, and his colleagues in the Epilepsy Center at Boston Children’s Hospital recently conducted the first human feasibility study on this topic. They sought to evaluate whether data derived solely from these wristbands could accurately predict various types of seizures in pediatric patients.
“For patients and their caretakers, the constant worry and uncertainty about presumably random patterns of seizure recurrence are among the most disabling aspects of epilepsy,” says Loddenkemper.
Going deeper with AI
Loddenkemper, the hospital’s director of clinical epilepsy research, and his team in the Loddenkemper Research Laboratory, recruited 69 epilepsy patients who were admitted to Boston Children’s long-term video-EEG monitoring unit between 2015 and 2018. These patients wore the Empatica wristband on either their wrist or ankle while in the unit. Using data collected from the wristbands, the researchers analyzed all types of seizures that occurred, including primary and secondary generalized and focal seizures.
To determine whether the wristbands could accurately predict seizures, they used a technique known as deep learning—a type of artificial intelligence, or AI, that relies on programmable neural networks to perform various classification tasks. Unlike traditional machine learning models, which still require some human guidance, deep learning algorithms use their own artificial neural networks to independently assess the accuracy of their predictions.
Forecasting seizures is feasible
Based on this work, Loddenkemper and his colleagues demonstrated that forecasting seizures is feasible using wristband data. Specifically, they found that the wristbands had a predictability that was better than chance in 43 percent of patients. The ability to predict seizures was independent of the type of seizure and the time of day it occurred, suggesting that this approach could benefit a wide range of epilepsy patients.
In addition, the prediction performance of the wristbands increased with additional patients and when they used all sensor modalities – a sign that the larger and richer the data collected, the more accurate the wristbands might be.
“Forecasting seizures with greater precision will enable us to titrate therapies to periods of greatest seizure susceptibility, with potentially fewer side effects at times of lower risk and better therapeutic coverage at times of greater risk,” Loddenkemper explains. Although the potential use of such wristbands is still several years away, this study represents an important first step in harnessing the power of wearables to improve patient care.
The study, led by former Boston Children’s researcher Christian Meisel, MD, Ph.D., now at the Universitätsmedizin Berlin and Berlin Institute of Health, was published in the October 11, 2020 issue of Epilepsia.
The unexpected nature of epileptic seizures represents the major clinical disability of epilepsy1. The mechanisms underlying the transition from a normal to a seizure state are currently an open question2–4. Unraveling the mechanisms underlying seizure generation could form the basis of much needed new treatment strategies, particularly for patients where existing treatments are ineffective.
Abrupt state changes in natural systems, including the onset of seizures, can, in principle, be due to critical transitions5. A characteristic of a system that is approaching a critical transition is a phenomenon called “critical slowing down”. Critical slowing down refers to the tendency of a system to take longer to return to equilibrium after perturbations, indicated by an increase in signal variance and autocorrelation.
Generally, critical slowing down can be expected if a system is driven towards the transition point at a moderate pace6 and if the basin of attraction around the equilibrium point can be approximated by linear-stability analysis7. It has been observed in many systems, including cell population collapse in bacterial cultures8 and crashes in financial markets9.
Critical transitions have been employed to describe neural systems, such as onset of depression10, pharmacologically induced cortical state changes11–13, onset of spiking in neurons14, and termination of epileptic seizures15.
It has been hypothesized that the rapid transition from normal brain activity to an epileptic seizure also corresponds to a critical transition3,4,16–19. However, empirical evidence for this hypothesis has been missing in humans, which may be due to lack of long-term recordings as well as intra-patient and inter-patient variability.
Empirical validation of critical slowing down in humans would provide vital support for current theoretical models of seizure generation and of the dynamics of the brain in general. Furthermore, it could aid in forecasting seizures and potential titration of epilepsy therapies.
Computational neural models are powerful tools for studying the dynamics of the brain. Numerous computational models of epilepsy suggest that seizures reflect a change in brain state via a critical transition16,18,20.
Mathematical analyses of dynamic systems, combined with simulations, enable classification of bifurcations and critical transitions21. Simulations enable controlled experiments that vary the parameters of the model and reveal statistical markers that are representative of transition susceptibility, such as increases in signal variance and autocorrelation.
While some methods have been developed to track control parameters (variables that drive changes in state) from clinically captured electroencephalography (EEG) in epilepsy22,23, this approach is not straightforward. Alternatively, tracking the statistical markers related to critical slowing down in clinical EEG recordings may constitute a direct test of the hypothesis that seizures occur via a critical transition.
In this paper, we test the hypothesis that markers of critical slowing down can be used as a biomarker of seizure susceptibility. We examine hallmark signals of critical slowing down using a continuous intracranial electroencephalography (iEEG) dataset from the first-in-human trial of an implanted seizure prediction device that was recorded over multiple years24.
As the markers of critical slowing down can potentially change over very long timescales, the long duration dataset used for this analysis provides a unique opportunity where critical slowing down in humans can be robustly investigated.
We show that the autocorrelation and variance of the iEEG signals are modulated by patient-specific cycles over long temporal scales. Furthermore, we show that modulations of the variance and autocorrelation are related to seizure susceptibility—a probabilistic propensity to have seizures.
reference link: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7195436/
More information: Christian Meisel et al. Machine learning from wristband sensor data for wearable, noninvasive seizure forecasting, Epilepsia (2020). DOI: 10.1111/epi.16719