Real-world Evaluation of an
FDA-cleared Wrist-worn Seizure Detector

Abstract number :1.191
Submission category : 2. Translational Research / 2A. Human Studies
Year : 2024
Submission ID : 1096
Source : www.aesnet.org

Presentation date : 12/7/2024 
Presenting Authors: 
Weixuan Vincent Chen, PhD – Empatica
Giulia Regalia, PhD – Empatica
Davide Cassani, PhD – Empatica
Rosalind Picard, ScD – MIT Media Lab
Marisa Cruz, MD – Empatica

Rationale: Artificial intelligence (AI) with multi-modal wearables has shown accurate detection of generalized tonic-clonic (GTCS) seizures [1], with prior evaluations limited to epilepsy monitoring unit (EMU) data [2]. Since the EMU limits a patient’s range of activity, we evaluated the performance of a wrist-worn FDA-cleared AI-based GTCS detection device in real-life settings, for both adult and pediatric patients.

[1] Donner E, Devinsky O, Friedman D. Wearable Digital Health Technology for Epilepsy. N Engl J Med. 2024;390(8):736-745.
[2] Onorati F, Regalia G, Caborni C, et al. Prospective Study of a Multimodal Convulsive Seizure Detection Wearable System on Pediatric and Adult Patients in the Epilepsy Monitoring Unit. Front Neurol. 2021;12:724904.

Methods: Patients (n=1,444) at risk of GTCS wore an FDA-cleared wrist-worn GTCS detection and alerting device configured with two alerting modes: High Sensitivity (HS) (default) and Low Sensitivity (LS) (optional, to minimize false alarms during daytime activities). When a likely GTCS is detected, the wristband triggers an alert to caregivers. At the time of the alert or afterward, the patient or the caregiver can confirm the event as a GTCS or mark it as a false alert. Missed GTCS can also be manually logged. “True GTCS events” were defined as either those detected by the device and manually confirmed by the patient, or those manually logged by the patient.

Data inclusion criteria were: (1) patients in the US or EU; (2) data recordings Jan 2020 - Dec 2023; (3) patient-months with wear-hours exceeding 22 days in 30 days; (4) patient-months with either 100% of device-triggered events reviewed by the user or with no device-triggered events. Rest periods were detected by an FDA-cleared actigraphy-based rest detection algorithm.

Endpoints evaluated were positive percentage agreement (PPA) and daily false alarm rate (FAR), with PPA as the ratio of true positive GTCS to total GTCS, corrected for multiple GTCS per patient, and overall FAR as the total false alarms divided by the total days (day = 24 hours of recorded data). Mean FAR = population average of individual false alarm rates. All calculations were retrospective, evaluating: (i) HS mode for all data, and (ii) LS mode for non-rest data. Confidence intervals (CI) used non-parametric bootstrapping.

Results: Table 1 summarizes the final analysis dataset, and Table 2 shows the evaluations for subgroups. HS mode detected 6112 out of 6346 GTCS, providing a corrected PPA=0.96 (CI_PPA_low= 0.95 overall, 0.95 adult, 0.93 pediatric). For patients with at least 2 seizures over the study period, 97.6% had a PPA >0.70, and over 85% had a PPA >0.99. The overall FAR was 0.75 (adult=0.69, pediatric=0.84), with a mean FAR=0.82. LS mode lowered the corrected PPA=0.80, (CI_PPA_low=0.71), but improved FAR (overall FAR=0.29, mean FAR=0.32).

Conclusions: GTCS detection using a wrist-wearable device exceeded FDA requirements [CI_PPA_low >70% (HS) or >60%(LS), and a FAR< 2 (HS) or < 1.5 (LS)], demonstrating accurate and clinically useful performance across daily-life activity.

Funding: Funded by Empatica