Cardiac-based detection of seizures in children with epilepsy

Highlights

• Commercial wearables can be used for real-time seizure detection to mitigate adverse events.

• Cardiac-based detection can be effective for seizures without motor activity.

• Adding movement-based features may improve detection time for some seizures.

• A patient-independent algorithm can detect generalized seizures with high sensitivity.

Abstract

Introduction

We evaluated a multi-parametric approach to seizure detection using cardiac and activity features to detect a wide range of seizures across different people using the same model.

Methods

Electrocardiogram (ECG) and accelerometer data were collected from a chest-worn sensor from 62 children aged 2–17 years undergoing video-electroencephalogram monitoring for clinical care. ECG data from 5 adults aged 31–48 years who experienced focal seizures were also analyzed from the PhysioNet database. A detection algorithm was developed based on a combination of multiple heart rhythm and motion parameters.

Results

Excluding patients with multiple seizures per hour and myoclonic jerks, 25 seizures were captured from 18 children. Using cardiac parameters only, 11/12 generalized seizures with clonic or tonic activity were detected as well as 7/13 focal seizures without generalization. Separately, cardiac parameters were evaluated using electrocardiogram data from 10 complex partial seizures in the PhysioNet database of which 7 were detected. False alarms averaged one per day. Movement-based parameters did not identify any seizures missed by cardiac parameters, but did improve detection time for 4 of the generalized seizures.

Conclusion

Our data suggest that cardiac measures can detect seizures with bilateral motor features with high sensitivity, while detection of focal seizures depends on seizure duration and localization and may require customization of parameter thresholds.

Introduction

Mortality risk within the population with epilepsy is 2–3 times higher than in a healthy population [1]. The ability to detect a seizure and alert a caregiver would enable timely intervention for mitigating adverse events including traumatic accidents and sudden unexpected death in epilepsy (SUDEP) which together account for over 60% of epilepsy-related deaths [1]. Seizure detection also enables accurate logging of seizures to support drug trials and patient medication management.

Electroencephalogram (EEG) monitoring is the gold standard method for detecting seizures in the clinical setting, but is not convenient for daily monitoring due to the requirement for electrodes on the head and a large volume of data to analyze. Non-EEG-based research devices and commercially available systems aimed at detecting seizures and warning caregivers have taken many different approaches [2], [3], [4]. Systems that identify abnormal movements related to convulsive seizures using sensors that are worn or placed on a mattress are the most prevalent. Movement-based systems can achieve high sensitivity (∼90%) for detecting generalized tonic-clonic seizures (GTCS), but are prone to false detections (1–10/day) [2], [5]. Improved specificity has been achieved using electrodermal activity (EDA), which reflects sympathetic regulation of sweat gland activity, in combination with movement-based detection. Detection with over 90% sensitivity for GTCS and a low false-positive rate (0.2–1/day) has been reported with this approach [6], [7] which is commercially available in the Empatica Embrace device.

Not all seizures involve shaking movements, but other physiological changes may manifest which may be leveraged for detection including autonomic and cardiovascular functions that can be monitored with unobtrusive sensors [3], [8], [9]. Systems leveraging these physiological changes may also be beneficial for SUDEP prevention because underlying arrhythmias and autonomic nervous system (ANS) dysfunction have been proposed as possible contributors to SUDEP [10]. A variety of methods have been reported to characterize cardiac changes during seizures. Algorithms using heartrate (HR) alone can achieve over 90% sensitivity, but tend to be error prone (i.e., over 1FP/hour) [3]. Algorithms using heart rate variability (HRV) measures can also achieve high sensitivity, but again false-positive rates are high (0.49–5.4/h) [3], [11]. Developing generalizable algorithms using cardiac data is challenging because cardiac profiles during both seizure and non-seizure periods can vary considerably between individuals and within-individuals for different types of seizures [12], [13], [14]. Adding patient specificity to algorithms improves performance [11], [15], [16], but can be difficult to implement as existing seizure data for each specific patient are required. A possible solution is the use of an initial patient-independent algorithm that is automatically adapted [17].

Another challenge of a real-time seizure detection based on cardiac activity is developing a fully automated algorithm that is robust to variations in signal quality. Approaches demonstrated on manually corrected HR data or evaluated only on short segments around seizures [13], [16] may not translate well to realistic monitoring applications. Additionally, some methods for quantifying ANS function are resource intensive, making real-time implementation on a mobile platform difficult. Consequently, few groups have demonstrated integrated systems [18], [19], [20], and these have only been validated with small numbers of users.

In this study, we present a real-time seizure algorithm that can be used with a variety of commercial sensors that provide accurate interbeat interval (IBI) data. The algorithm was developed with data from a single-lead, chest-worn electrocardiogram (ECG) sensor with an integrated accelerometer. Seizures were detected using a combination of simple, patient-independent HR and HRV indicators of elevated autonomic activity. Motion-related parameters were also explored. We evaluated algorithm performance across a range of seizure types and achieved high sensitivity with a low false-positive rate. We implemented the algorithm on a custom standalone unit that received data streamed from the chest-worn ECG sensor in real-time. This prototype unit could be replaced by a smartphone in future applications.