Abstracts

Rationale: Wearable devices (WD) could solve the problem of long-term focal seizure detection at home. Nonetheless, most published studies had a short duration and addressed mainly generalized tonic-clonic seizures. We studied the diagnostic yield of a WD, the Sensor Dot, in focal seizures detection at home for 3 months.

Methods: Adults with focal refractory epilepsy used the WD at least 16 hours/day for 90 days and reported their seizures via an electronic seizure diary. The WD recorded two electroencephalography (EEG) channels through dry electrode patches placed behind the ears (BTE) on the mastoid bone. Data was transmitted to a secure cloud via Wi-Fi (Figure 1). The files were visually inspected offline. Segments with poor signal quality were discarded, but those with movement artefacts during daily activities were kept. A patient-independent seizure-detection algorithm flagged possible seizures. This algorithm was trained with EEG data acquired with cup electrodes BTE on a different population. Patients’ reported events were the gold standard. Taking into consideration class imbalance, the following metrics evaluated the algorithm’s performance:_x000D_ _x000D_ Recall: (TP)/(TP + FN), where TP = true positives and FN = false negatives_x000D_ False positives (FP) per hour (FP/hour): FP / length of the recording_x000D_ Precision: (TP)/(TP+FP)_x000D_ F1-score: 2*(Precision * Recall)/(Precision + Recall)_x000D_ Detection delay: time interval between EEG seizure onset and detection by the algorithm_x000D_
Results: Eleven patients (7 males, median age of 39 years) participated in the study (Table 1). After visual inspection for signal quality, 29.4% of the data were analyzed with the algorithm, which corresponded to 6481.85 hours of measurements. The review time for 24 hours of flagged EEG was 9.15 minutes [standard deviation [SD], 0.2 minutes]. The mean values for recall and precision were 16.44% [SD, 21.68%] and 0.04% [SD, 0.05%]. The average FP/hour was 7.09 [SD, 1.61]. When only considering the reported seizures that were actually visible on the EEG, the mean recall increased to 42% [SD, 29%], without a significant improvement in the F1 score due to the high amount of FP/hour. The median detection delay was 2.5 seconds [IR, 1-9.5]. Thirteen seizures not reported by the patients were detected by the algorithm. All occurred during sleep. The main reasons for the high FP/h rate were movement artefacts. The low and variable recall among different patients might be related to differences in the evolution of the EEG pattern during the seizures and the contamination of the EEG signals by movement artefacts. 

Conclusions: A two-channel scalp EEG WD is able to detect focal seizures in patients with refractory epilepsy during long term monitoring at home. Nevertheless, the current setup and workflow limit its use in clinical practice. Further research must be focused on artefact removal and patient-specific algorithms, as well as the integration of other biosignals. New materials (e.g., hydrogel electrodes) and improved patient training might increase the signal quality._x000D_
Funding: The study was supported by a Personalized Medicine Interdisciplinair Coöperatief Onderzoek (ICON, HBC.2019.2521) project of Vlaams Agentschap Innoveren en Ondernemen (VLAIO).