Abstracts

Rationale: Embrace is the first FDA approved wrist-worn device combining Accelerometer (ACM) and Electrodermal Activity (EDA) to detect and alert to generalized tonic-clonic seizures (GTCSs). A multi-site clinical evaluation of the machine learning classifier embedded in Embrace validated 94.5% sensitivity of the GTCS detector with a false alarm rate typically lower than 0.2/day (Onorati et al, Epilepsia 2017). When applied on real-life data, Embrace yielded comparable performances in different analyses (Caborni et al, 32^nd International Epilepsy Congress 2017; Regalia et al, American Epilepsy Society meeting 2017; Onorati et al, 12^th European Congress on Epileptology 2018). The ankle has been reported as more comfortable than the wrist by a significant portion of patients (Fedor et al, 54th Annual Meeting of the Society for Psychophysiological Research, 2014), especially for pediatric patients. The aim of this work is to investigate whether the Embrace detection algorithm, designed for wrist data, shows good generalizability to ankle data gathered in clinical settings. Methods: Data were collected during clinical v-EEG monitoring and consist of 191 recordings taken from 86 patients (81 children, average age: 8 y; 5 adults, average age: 22 y) wearing an Empatica E4 device, provided with EDA and 3­axis ACM sensors, at the ankle. Seizures were labeled by two independent epileptologists. EDA and ACM recordings were analyzed off­line with the detection algorithm, previously trained only on wrist data. Sensitivity (Se) was computed as the ratio of GTCSs that triggered an alert. False alarm rate (FAR) was computed as the number of alerts not corresponding to GTCS events, divided by the total recorded hours, normalized by 24 hrs. Results: Overall, 150 days of data were recorded, including 15 generalized GTCSs from 10 patients (9 children). 14 GTCSs were successfully detected (Se=93.3%) by the algorithm. The missed seizure (from a 6 y/o female) exhibited a milder ACM pattern. The cumulative FAR was 0.03, roughly corresponding to 1 false alarm per month. Seventy-three patients (85%) did not experience false alarms; 9 patients (10%) had a FAR between 0 and 1; 2 patients (2.5%) had a FAR between 1 and 2, while 2 patients had a FAR higher than 2. Conclusions: The performance of an ACM and EDA-based classifier trained on wrist data was evaluated on ankle data for the first time. Both the sensitivity and the FAR indicate that the classifier is able to generalize well to motor and autonomic seizures patterns sensed at the lower limbs. The much lower FAR observed on ankle data may be due to the lower occurrence of seizure-like patterns at the ankle. Future work will evaluate the performance also in more challenging real-life ambulatory environments. Funding: No fund