Detection of absence seizures using a glasses-type eye tracker

Childhood absence epilepsy (CAE) is the most common form of childhood epilepsy, accounting for 8–15% of all childhood-onset epilepsy cases (Tenney and Glauser, 2013). CAE patients are generally unable to report their own seizure events, and the risk of seizure-related injuries is estimated to be 3% per person-year (Tenney and Glauser, 2013). Wearable devices accurately detecting absence seizures are eagerly awaited for monitoring the rate of seizure occurrence at home. Ictal forced upward eye deviation is one of the objective symptoms of CAE. A video-electroencephalography (EEG) study reported that CAE patients had an eye deviation in up to 80% of seizure events (Tenney and Glauser, 2013). Seizure-related clinical signs, including upward eye deviation, are suggested to occur within 1.0 s after the ictal EEG onset (Tenney and Glauser, 2013). A mobile glasses-type eye tracker is useful to continuously monitor eye movement in real-world settings (Simpson et al., 2019). This methodological study tested the feasibility of using a glasses-type eye tracker for seizure detection. We determined how accurately and rapidly the sustained upward eye deviation observed by the eye tracker would detect the electrographic onset of absence seizure events marked by two board-certified epileptologists.

Why do we need to establish a non-EEG seizure detection device? A scalp EEG-based algorithm was reported to achieve a sensitivity of 97.2%, no false detection, and a detection latency of 0.74 s in a study of 125 absence seizures occurring over 11.5 h of recording time in 20 patients (Ulate-Campos et al., 2016). A previous survey reported that only 13–20% of patients with drug-resistant focal epilepsy could imagine wearing scalp EEG electrodes on a long-term basis (Ulate-Campos et al., 2016). Then, why not using a video-based detection method? A study of five CAE patients reported that a video-based algorithm detected all 21 absence seizures that took place in a total period of 9.5 min (Sathyanarayana et al., 2015). The video-based algorithm was effectively designed to identify a staring episode lasting longer than 3 s, but the false alarm rate (FAR) was not clarified in that study. It may be challenging to capture video footage of the face continuously in real-world settings. Thus, we expected that a mobile glasses-type eye-tracker would serve as a viable wearable seizure detection device.

A seven-year-old boy with CAE simultaneously underwent eye-tracking and video-EEG monitoring. We obtained written informed consent from the legal parent. The Wayne State University Institutional Review Board has approved the present study. We provided details of the analysis method in the Supplementary Material.

We identified nine absence seizures during the 13.5-minute recoding. Anti-epileptic therapy was initiated afterward. The median seizure duration was 11.4 s (standard deviation: ±2.4). Eight of the nine seizures (88.9%) were associated with an upward eye deviation, whereas the remaining one did not show an eye deviation. Video S1 shows the temporal relationship between the ictal discharge and eye movements in a seizure event (Fig. 1).

Figure 1. Ictal eye movement and electroencephalography (EEG).

The figure provides a snapshot of Video S1, which demonstrates the temporal relationship between the ictal EEG discharge at Cz and eye movement during the eighth seizure event. Red circles: vertical eye positions (positive pixel: upward gaze). Blue circles: horizontal eye positions (positive pixel: rightward gaze). Two seconds after the ictal EEG onset, the patient showed an upward/leftward eye deviation. The eye position came back to the center at around the ictal offset. *: eye blinks. AV: common average reference. Rt: right. Lt: left.

With an analysis epoch (T) of 4 s, the eye-tracker achieved the optimal seizure detection performance characterized by a sensitivity of 88.9%, FAR of 0, and detection latency of 2.87 s (Supplementary Table S1). None of the tested settings detected a seizure event not accompanied by an eye deviation. Increasing T to 6 or 8 undesirably prolonged the detection latency without improving the sensitivity or FAR. Decreasing T to 2 reduced the detection latency to 1.53 s but increased the FAR to 4 per hour.

To our best knowledge, this is the first-ever study that assesses the feasibility of using a glasses-type eye-tracker for detection of absence seizures. Our simple algorithm was capable of detecting 88.9% of seizure events without a false alarm and within 3 s after the ictal EEG onset. Mobile glasses-type eye trackers have the potential to serve as a viable wearable seizure detection device and may provide complementary information to other seizure diary tools. Caregivers may visually identify substantial proportions of the seizure events occurring in front of them, but continuous visual monitoring requires enormous cognitive efforts. Up to 80% of CAE patients are suggested to have seizures characterized by sustained upward eye deviation (Tenney and Glauser, 2013). Such ictal forced eye deviation is attributed to pathological excitation of the bilateral frontal eye fields (Bajwa et al., 2006). When looking upward in daily life, conversely, healthy humans swiftly return the eyes to an immediately-prior, less-eccentric, more-central position with a coordinating head movement.

Many investigators have utilized the mobile glasses-type eye tracker in real-world settings to investigate the attention distribution (Simpson et al., 2019). We suggest that eye trackers have the potential to provide seizure semiology information and improve the accuracy of other seizure detection methods at home or school. However, some may argue that it is too optimistic to expect the eye tracker to optimally function as a stand-alone seizure detection tool. Our eye-tracking system is insensitive to seizures without eye deviation, whereas an EEG-based algorithm was reported to show a sensitivity of 97.2% (Ulate-Campos et al., 2016). With an analysis epoch (T) of 4 s, our eye tracker detected seizure events with a mean latency of 2.87 s. In contrast, the EEG-based algorithm was reported to detect seizures with a latency of 0.74 s (Ulate-Campos et al., 2016). With T of 2 s, the eye tracker had a mean latency of 1.53 s but suffered from a FAR of 4 per hour. It may be infeasible to expect the eye tracker to have a sub-second detection latency since the discernible clinical changes occur 1.0 s after the EEG onset in CAE patients (Tenney and Glauser, 2013).

This study included a single patient who had nine seizures during a short video-EEG recording period. Thus, we must interpret the observed FAR with caution; a low FAR could be attributed to the relatively short interictal state. Further studies of large sample sizes for more extended periods are needed to assess the generalizability of the results of this preliminary study and are expected to clarify the engineering issues to be addressed for optimal device operation.

Abbreviations

CAE

childhood absence epilepsy

EEG

electroencephalography

FAR

false alarm rate