Seizure forecasting: Bifurcations in the long and winding road

Comparison between epileptic seizure prediction and forecasting based on machine learning

Epilepsy affects around 1% of the population worldwide. Anti-epileptic drugs are an excellent option for controlling seizure occurrence but do not work for around one-third of patients. Warning devices employing seizure prediction or forecasting algorithms could bring patients new-found comfort and quality of life. These algorithms would attempt to detect a seizure's preictal period, a transitional moment between regular brain activity and the seizure, and relay this information to the user. Over the years, many seizure prediction studies using Electroencephalogram-based methodologies have been developed, triggering an alarm when detecting the preictal period. Recent studies have suggested a shift in view from prediction to forecasting. Seizure forecasting takes a probabilistic approach to the problem in question instead of the crisp approach of seizure prediction. In this field of study, the triggered alarm to symbolize the detection of a preictal period is substituted by a constant risk assessment analysis. The present work aims to explore methodologies capable of seizure forecasting and establish a comparison with seizure prediction results. Using 40 patients from the EPILEPSIAE database, we developed several patient-specific prediction and forecasting algorithms with different classifiers (a Logistic Regression, a 15 Support Vector Machines ensemble, and a 15 Shallow Neural Networks ensemble). Results show an increase of the seizure sensitivity in forecasting relative to prediction of up to 146% and in the number of patients that displayed an improvement over chance of up to 300%. These results suggest that a seizure forecasting methodology may be more suitable for seizure warning devices than a seizure prediction one.

Seizure forecasting: Where do we stand?

A lot of mileage has been made recently on the long and winding road toward seizure forecasting. Here we briefly review some selected milestones passed along the way, which were discussed at the International Conference for Technology and Analysis of Seizures-ICTALS 2022-convened at the University of Bern, Switzerland. Major impetus was gained recently from wearable and implantable devices that record not only electroencephalography, but also data on motor behavior, acoustic signals, and various signals of the autonomic nervous system. This multimodal monitoring can be performed for ultralong timescales covering months or years. Accordingly, features and metrics extracted from these data now assess seizure dynamics with a greater degree of completeness. Most prominently, this has allowed the confirmation of the long-suspected cyclical nature of interictal epileptiform activity, seizure risk, and seizures. The timescales cover daily, multi-day, and yearly cycles. Progress has also been fueled by approaches originating from the interdisciplinary field of network science. Considering epilepsy as a large-scale network disorder yielded novel perspectives on the pre-ictal dynamics of the evolving epileptic brain. In addition to discrete predictions that a seizure will take place in a specified prediction horizon, the community broadened the scope to probabilistic forecasts of a seizure risk evolving continuously in time. This shift of gears triggered the incorporation of additional metrics to quantify the performance of forecasting algorithms, which should be compared to the chance performance of constrained stochastic null models. An imminent task of utmost importance is to find optimal ways to communicate the output of seizure-forecasting algorithms to patients, caretakers, and clinicians, so that they can have socioeconomic impact and improve patients' well-being.

Learning to generalize seizure forecasts

OBJECTIVE

Epilepsy is characterized by spontaneous seizures that recur at unexpected times. Nonetheless, using years-long electroencephalographic (EEG) recordings, we previously found that patient-reported seizures consistently occur when interictal epileptiform activity (IEA) cyclically builds up over days. This multidien (multiday) interictal-ictal relationship, which is shared across patients, may bear phasic information for forecasting seizures, even if individual patterns of seizure timing are unknown. To test this rigorously in a large retrospective dataset, we pretrained algorithms on data recorded from a group of patients, and forecasted seizures in other, previously unseen patients.

METHODS

We used retrospective long-term data from participants (N = 159) in the RNS System clinical trials, including intracranial EEG recordings (icEEG), and from two participants in the UNEEG Medical clinical trial of a subscalp EEG system (sqEEG). Based on IEA detections, we extracted instantaneous multidien phases and trained generalized linear models (GLMs) and recurrent neural networks (RNNs) to forecast the probability of seizure occurrence at a 24-h horizon.

RESULTS

With GLMs and RNNs, seizures could be forecasted above chance in 79% and 81% of previously unseen subjects with a median discrimination of area under the curve (AUC) = .70 and .69 and median Brier skill score (BSS) = .07 and .08. In direct comparison, individualized models had similar median performance (AUC = .67, BSS = .08), but for fewer subjects (60%). Moreover, calibration of pretrained models could be maintained to accommodate different seizure rates across subjects.

SIGNIFICANCE

Our findings suggest that seizure forecasting based on multidien cycles of IEA can generalize across patients, and may drastically reduce the amount of data needed to issue forecasts for individuals who recently started collecting chronic EEG data. In addition, we show that this generalization is independent of the method used to record seizures (patient-reported vs. electrographic) or IEA (icEEG vs. sqEEG).

Ultrafast review of ambulatory EEGs with deep learning

OBJECTIVE

Interictal epileptiform discharges (IED) are hallmark biomarkers of epilepsy which are typically detected through visual analysis. Deep learning has shown potential in automating IED detection, which could reduce the burden of visual analysis in clinical practice. This is particularly relevant for ambulatory electroencephalograms (EEGs), as these entail longer review times.

METHODS

We applied a previously trained neural network to an independent dataset of 100 ambulatory EEGs (average duration 20.6 h). From these, 42 EEGs contained IEDs, 25 were abnormal without IEDs and 33 were normal. The algorithm flagged 2 second epochs that it considered IEDs. The EEGs were provided to an expert, who used NeuroCenter EEG to review the recordings. The expert concluded if each recording contained IEDs, and was timed during the process.

RESULTS

The conclusion of the reviewer was the same as the EEG report in 97% of the recordings. Three EEGs contained IEDs that were not detected based on the flagged epochs. Review time for the 100 EEGs was approximately 4 h, with half of the recordings taking <2 minutes to review.

CONCLUSIONS

Our network can be used to reduce time spent on visual analysis in the clinic by 50-75 times with high reliability.

SIGNIFICANCE

Given the large time reduction potential and high success rate, this algorithm can be used in the clinic to aid in visual analysis.

Removing artefacts and periodically retraining improve performance of neural network-based seizure prediction models

The development of seizure prediction models is often based on long-term scalp electroencephalograms (EEGs) since they capture brain electrical activity, are non-invasive, and come at a relatively low-cost. However, they suffer from major shortcomings. First, long-term EEG is usually highly contaminated with artefacts. Second, changes in the EEG signal over long intervals, known as concept drift, are often neglected. We evaluate the influence of these problems on deep neural networks using EEG time series and on shallow neural networks using widely-used EEG features. Our patient-specific prediction models were tested in 1577 hours of continuous EEG, containing 91 seizures from 41 patients with temporal lobe epilepsy who were undergoing pre-surgical monitoring. Our results showed that cleaning EEG data, using a previously developed artefact removal method based on deep convolutional neural networks, improved prediction performance. We also found that retraining the models over time reduced false predictions. Furthermore, the results show that although deep neural networks processing EEG time series are less susceptible to false alarms, they may need more data to surpass feature-based methods. These findings highlight the importance of robust data denoising and periodic adaptation of seizure prediction models.

Decrease in wearable-based nocturnal sleep efficiency precedes epileptic seizures

Introduction

While it is known that poor sleep is a seizure precipitant, this association remains poorly quantified. This study investigated whether seizures are preceded by significant changes in sleep efficiency as measured by a wearable equipped with an electrocardiogram, respiratory bands, and an accelerometer.

Methods

Nocturnal recordings from 47 people with epilepsy hospitalized at our epilepsy monitoring unit were analyzed (304 nights). Sleep metrics during nights followed by epileptic seizures (24 h post-awakening) were compared to those of nights which were not.

Results

Lower sleep efficiency (percentage of sleep during the night) was found in the nights preceding seizure days (p < 0.05). Each standard deviation decrease in sleep efficiency and increase in wake after sleep onset was respectively associated with a 1.25-fold (95 % CI: 1.05 to 1.42, p < 0.05) and 1.49-fold (95 % CI: 1.17 to 1.92, p < 0.01) increased odds of seizure occurrence the following day. Furthermore, nocturnal seizures were associated with significantly lower sleep efficiency and higher wake after sleep onset (p < 0.05), as well as increased odds of seizure occurrence following wake (OR: 5.86, 95 % CI: 2.99 to 11.77, p < 0.001).

Discussion

Findings indicate lower sleep efficiency during nights preceding seizures, suggesting that wearable sensors could be promising tools for sleep-based seizure-day forecasting in people with epilepsy.