Detecting epileptic seizures with multimodal non-EEG data from wearables

Epilepsy is one of the most prevalent chronic neurological disorders, affecting millions worldwide throughout all societal groups. Epilepsy manifests in those affected as reoccurring seizures with a wide range of different symptoms at variable intervals and severity. The current gold standard to diagnose and monitor epilepsy is video-electroencephalography. Patients with epilepsy visit monitoring units for a few days, and clinicians provoke seizures through various means, hoping to get enough information for precise diagnosis and treatment. However, this procedure is not viable during the patients' daily lives and over more extended periods. Furthermore, the handwritten diaries that some patients keep have proven unreliable, typically severely under-counting the number of seizures occurring. An alternative for ultra-long-term monitoring is needed to improve current treatments and facilitate the development of new therapy options. This thesis investigates the potential of multimodal non-electroencephalography data recorded from wearable devices as a tool for seizure detection in the context of automated seizure diaries. It furthermore explores a potential application of seizure detection in the context of an automatic alarm system.

The work featured in this thesis produces and employs a new data set of wearable biosignal data, recorded at two European epilepsy centers in the context of a European collaborative research project. Over 200 patients with epilepsy were recruited at the two epilepsy monitoring units, and over 300 epileptic seizures of varying types were recorded with a wearable device. Here, the Empatica E4 is used, a research-grade wrist-worn wearable that captures the biosignal modalities of accelerometry (movement), electrodermal activity (electrical skin conductance), and blood volume pulse (optical pulse measurement via photoplethysmography). This data set was the basis for several data analysis studies concerning the evaluation of seizure detection methodologies.

This thesis compiles and provides a framework for several contributions of the author concerning the detection of epileptic motor seizures with multimodal non-electroencephalography data from wearables. Specifically, the included studies investigated those seizures with movement manifestations in the limbs and found detection systems based on supervised ensemble machine learning using physiological biosignal data to be viable.

One central part of this thesis is focused on convulsive tonic-clonic seizures, severe and dangerous seizures that start in or progress to both hemispheres of the brain. During these seizures, the awareness and consciousness of the affected patient are impaired, and high-amplitude, high-frequency jerks of the limbs and whole body occur. One of the studies presented here assessed an automatic detection methodology based on a combination of accelerometry and electrodermal activity signals. A supervised ensemble machine learning model is trained on expert-labeled data and evaluated on an out-of-sample test set. It performs at least on par with state-of-the-art related work, correctly classifying more than 90 percent of seizure events with false alarm rates of less than 0.5 per day. The suggested methodology performs better than the average monomodal detection system in related work.
Convulsive tonic-clonic seizures are typically followed by a period of unconsciousness and immobility, significantly increasing the risk of sudden unexpected death in epilepsy. A further study investigates the utility of wearable biosignal data to detect and gauge this period based on a heuristic detection using accelerometry signals. Contingent on a prior automatic detection of the seizure, the methodology was able to classify all instances of immobility in the data set correctly.

Another essential segment of this thesis highlights the detection of focal seizures with data from wearables. Focal seizures typically have very heterogeneous symptoms when regarded across patients. They include body movements of different kinds, responses of the autonomic nervous system, and psychological indications. The research included here analyzed only those focal seizures with specific movements of the limbs. An early exploratory study investigated the impact of the high variance of focal motor seizures on biosignals and the performance of seizure detection based on those signals.
An additional study then considered individualized and generic models for detecting focal motor seizures based on the biosignals recorded by the wearable. The study found the optically measured blood volume pulse data to be highly impacted by noise from motion artifacts. Furthermore, generic models performed considerably worse than those specific to an individual patient, with high false alarm rates. Thus, for focal seizure detection, custom-made detection models for individual patients are likely to be the most robust methodology, and are specifically suitable for a subset of patients with epilepsy who experience characteristic seizures.

This thesis concludes that while generic seizure detection models may be sufficient for highly convulsive seizures and under in-hospital conditions, they are currently not feasible for detecting focal seizures with fewer or no movements. Conversely, patient-specific detection methodologies are promising for non-convulsive motor seizures. Detection models that individualize over time may eventually become the best option for ultra-long-term seizure detection. Specifically, the included studies investigated detection systems based on supervised ensemble machine learning using physiological biosignal data. Results showed them to be feasible for detecting convulsive and less-convulsive seizures with manifestations including movements of the limbs.