Absence seizures are the main seizure type in various idiopathic generalized epilepsy syndromes, among others, in juvenile absence epilepsy (JAE) (Hirch et al., 2022). Typical absence seizures are characterized by motor arrest and impaired awareness. On the electroencephalogram (EEG), they are visualized by 3–4.5 Hz generalized spike-and-wave discharges. In patients with JAE, lifelong treatment with antiseizure medications may be required as relapse after withdrawal is high (83%) (Hirch et al., 2022). Even patients who are considered to be seizure free under treatment, can still relapse (25%).

It is, therefore, paramount to accurately detect absence seizures in order to improve therapeutic decisions. However, the subtle clinical manifestation of absence seizures hampers the patient’s and the caregiver’s ability to correctly report the seizure in the seizure diary. In studies with patients with absence seizures, 86%–94% of the absences are not reported (Keilson et al., 1987; Akman et al., 2009; Swinnen et al., 2021). As such, the seizure diary is an unreliable method, although commonly used in clinical practice as well as surrogate endpoint in trials for antiseizure medication (Fisher et al., 2012).

Current diagnostics are based on the clinical history and in-hospital monitoring of seizures with video-EEG (which is considered the gold-standard for seizure detection). Hence, devices that could record the seizures during normal everyday life conditions could facilitate a better logging of the seizure events, as the use of ambulatory EEG is limited due to the complexity of placement and the visibility and constraints in daily life.

An explosion in the capability of monitoring individuals via sensors built into smartphones or wearable devices has been taking place. Following the views of the patients, the devices need to be “as ‘normal’ as possible,” “non-stigmatizing” and “non-intrusive” (Bruno et al., 2018). Furthermore, the majority of the patients were willing to use a wearable device for long-term seizure tracking, envisaged by the possibility of understanding their condition better. The detection of absence seizures with wearable devices is a very suitable use-case, since the number and location of electrodes is likely to be less important than in focal epilepsies (Duun-Henriksen et al., 2012).

However, the clinical adoption of wearable EEG remains challenging, since no phase-4 clinical study, proving the feasibility and comfort of the device in a home environment, has yet been conducted. Furthermore, there are challenges regarding the signal quality and the time needed by the neurologists to annotate the data. Hence, research for automated seizure detection algorithms, tailored for wearable EEG, is still in its infancy, especially in wearable detection of absence seizures (Kjaer et al., 2017; Dan et al., 2020; Swinnen et al., 2021; Japaridze et al., 2022), (semi) automated seizure detection would facilitate the adoption of EEG wearables in clinical practice. Furthermore, even the annotation of wearable EEG data, compared to vEEG, is not trivial (Zhang et al., 2022).

In a previous paper (Swinnen et al., 2021), we reported the performance of an EEG-based wearable, CE marked device to detect absence seizures, the Sensor Dot (SD; Byteflies). SD is a discrete, user-friendly wearable that consists of four electrodes, creating two ipsilateral channels. This device was developed during SeizeIT1 (2016–2019). The initial results currently reported belong to an extension of SeizeIT1, SeizeIT2 which is a larger multicenter trial with focus on clinical validation of the SD in people with typical absence, focal impaired awareness, and generalized tonic–clonic seizures (EIT Health: SeizeIT2, clinicaltrials.gov: NCT04284072) (EIT Health, 2020).

In the same paper we proposed a novel automated absence detection algorithm which achieved a mean sensitivity of 0.983 in the detection of absence seizures, with mean False Alarm (FA) per hour equal to 0.9138. The use of the automated algorithm reduced the review time of a 24 h recording from 1 to 2 h to around 5–10 min for the neurologists.

Nevertheless, the reported results in (Swinnen et al., 2021) were obtained from data acquired within the hospital environment. We argued that the first phase of the recording of each patient will consist of routine monitoring in the EMU after which these data can be used to train the algorithm. The obtained classifier can be used to detect seizures from the data acquired when the patient wears the SD at home. Similar assumptions are being made by all of the aforementioned studies with wearable devices and absence seizure detection. Even in the phase-3 clinical trial of Epihunter (Japaridze et al., 2022), the data were not obtained in an actual home environment [as proposed in (Beniczy and Ryvlin, 2018)].

As observed from the data obtained from the patients in the home environment with SD, the performance of the automated seizure annotation was lower, compared to the performance of the algorithm trained and tested with data obtained in the hospital (Swinnen et al., 2021), especially in terms of FAs. In the current paper we will present a pipeline for the automated absence seizure detection for in-home monitoring, which will be tested and validated, for the first time, with data obtained from patients with absence seizures in their home environment. The focus of the current paper will only be the algorithmic pipeline, used to flag the possible seizures, the physiological findings of the home studies will be released in future publications. The goal of the pipeline is to be used by the neurologists in the rest of the subjects and reduce the time needed to annotate the data. The neurologists are requested instead of reviewing the whole recording to only review the provided alarms from the algorithm.

In this section we will provide some initial results from the two home studies. The selection of the recordings of the patients that were analyzed was based on the number of seizures. We selected the first four patients of SeizeIt2 recorded in 2022 that had at least one absence seizure and the first 6 days of the first absence patient of PnP study for which at least one seizure was recorded. In both datasets, the onset and end of all 3-Hz generalized spike- and- wave discharges (SWD), lasting 3 s or more ¹ , were marked by neurologists. We will refer to these annotations as “absences,” these annotations were used as the ground truth for the performance metrics that will be presented.

In this section we will try to justify and explain the usability of the different blocks of our pipeline. We will compare the results with the three different training options and to the algorithm presented in (Swinnen et al., 2021) that was tested only with in-hospital recordings. We must note that for the PnP patient no in-hospital recording exists, hence only the first training option will be used. For the SeizeIt2 patients since we have a single day of in-home recordings we will use options two and three for training.

The following metrics were applied to determine the performance of the seizure detection algorithm [following the definitions in Vandecasteele et al. (2020)]:

• Sensitivity: T P T P + F N . A seizure was detected correctly (TP) if the detection occurred between the EEG onset and end of the seizure. A seizure was not detected (FN) if no overlap between the alarm and the seizure existed.

We will not report specificity measures since for all cases the specificity was above 99%, due to the highly imbalanced classes.

As can be noted from Table 2 in the PnP patient, the postprocessing decreases significantly the Fas/h, with a slight decrease of the sensitivity in most of the recordings. The removal of the bad segments (column 4) retains the low FA rate but increases the Sensitivity, reaching 1 in almost all of the days. There is minor influence from the postprocessing with the IMU data.

For the SeizeIt2 data (Table 3) we can see that the postprocessing has similar effect as in the PnP patient. We shall note the tremendous decrease in the FAs/h for Subj-3. Furthermore, the IMU postprocessing results in a slight decrease in the sensitivity in Subj-3. The decrease is due to the activity noticed in accelerometer and gyroscope during two seizures (Figure 5). We note the very low performance of all the versions of the algorithm in the last Subject.

When we train the SVM with data obtained from home recordings of other patients and “personalize” it with the hospital recording of the test patient (Option 3- Figure 4), we note that the performance in Subj-4 increases significantly and is comparable to the other Subjects (Table 4). There is a minor effect to the rest of the Subjects.

We must note that both for the PnP subject and the SeizeIt2 subjects, the algorithm found some seizures that were missed by the neurologists (and are currently counted as FAs, since they were not included in the gold standard). More specifically for the PnP subject eight seizures were found that were initially missed by the neurologist (the extra seizures per recording can be viewed in the last column of Table 2), while for the SeizeIt2 subjects 1 seizure was found (examples of missed seizures can be viewed in Figure 6.

According to the standards for testing and clinical validation of seizure detection devices (Beniczy and Ryvlin, 2018), our studies can be classified as a phase-4 study, since the accuracy and usability of the SD in a home environment is investigated in a multicenter trial (EIT Health: SeizeIT2, clinicaltrials.gov: NCT04284072). To the best of our knowledge, these are the first (preliminary) results of a phase-4 trial of an automated detection of absence seizures with a wearable device. As mentioned in the introduction the full analysis and report for the findings of the study will follow in another publication and the main focus of this paper is to investigate design choices in the pipeline used for the automatic detection of the seizure events.

The current state-of-the art have failed to provide accurate detection of the absence seizures with the home data mainly due to the high FA rate obtained. We have exhibited that our pipeline provides accurate detection of the absence seizures with very high sensitivity and low FAs/h. Furthermore, it managed to detect seizures that were initially missed by the neurologists. We shall note the difference in the amount of seizures missed by the two neurologists in the different studies which is probably due to the multiple recordings in the span of several months for the PnP study. The large amount of EEG data requires extended time for annotation of the seizures, hence, lower performance might be expected by the annotators. The need of automated seizure detection pipelines is evident when large amount of data are acquired. The lower signal quality of the in-home recordings and the presence of artifacts that did not exist in the hospital recordings, dictate the use of alternative blocks in our pipeline.

Furthermore, we have exhibited for the first time the added value of multimodality in the context of absence seizure detection. Multimodal seizure detection has been widely used for the detection of motor seizures (Leijten, 2017) and focal seizures (Vandecasteele et al., 2021). However, multimodal approaches have not been employed for the detection of absence seizures, mainly due to lack of movement or heart rate variability during typical absence seizures. We have shown that detection of movement with SD can help into distinguishing seizure events from movement and dramatically decrease FAs (Subj-3, SeizeIt2).

The thresholding based on the standard deviation of the IMU signals was adopted in order to decrease the significant amount of FAs that were caused by the movement of everyday life. The hospital data do not include such movement segments, hence, the classifier, when trained only with hospital data do not recognize those segments as background but as seizure segments. “Turning off” the IMU postprocessing can result in a detection of a very limited number of seizures, similar to the one presented in Figure 5, for which the standard deviation (std) of the IMU data is above the threshold.

The current pipeline uses IMU modality only for postprocessing, an initial attempt to use an hierarchical clustering for detection of walking and running artifacts had inferior performance to the simple threshold-based postprocessing. However, different “true” fusion approaches (Chatzichristos et al., 2022) either early or late ones will be further explored.

Concerning the different training approaches that we have proposed we can note that it seems that in some cases (Subj-4, SeizeIt2) the in-hospital data differ a lot from the in-home data. In such cases the addition to the training set of data from in-home recordings of other patients can be beneficial. When multiple days of in-home recordings exist the training of the model with labelled data from one of the days provides the best performance. We recommend the use of the first option of the pipeline if there are multiple day recordings per subjects and the third pipeline when a single day of recording exist per subject.

Further work will aim to independently replicate those results with the data collected from the different centers that participate in SeizeIt2 and with the rest of the patients from UZ Leuven. Correct interpretation of artifacts in SD recordings will also be helpful in the detection of other seizure types and reduction of FA in the home setting.

Furthermore, when the amount of the data collected allows, we will study approaches based on Neural Networks which have shown significant success in several seizure detection applications (Chatzichristos et al., 2020; Japaridze et al., 2022), very recently even with the detection of absence seizures with two-channel EEG (Hartmann et al., 2022).