Real-time seizure detection performance with Embrace alert
system: one year real-life setting case study
G. Regalia1, C. Caborni1, M. Migliorini1, F. Onorati1, R. W. Picard1,2
1 Empatica Inc, Milan, Italy
2 MIT Media Lab, Massachusetts Institute of Technology, Cambridge, Massachusetts, U.S.A
1st International congress on mobile health devices and seizure detection in epilepsy - Copenhagen 7-8 July 2017
•Automated mobile seizure detection devices can prompt caregivers
to intervene and provide objective seizure counts in daily life.
However, their performance outside epilepsy monitoring units (EMUs)
is largely unknown.
•The Embrace system is the first commercially available multimodal
system combining accelerometer and electrodermal activity to detect
and alert to convulsive seizures (CSs) (Figure 1).
•An automated machine-learning classifier was trained using video-
EEG labeled EMU data [1-3] and optimized using data from
outpatient settings .In aprior 3-month home study with one
patient and real-time detection and alerts, Embrace alerted to CS’s
with 92%sensitivity and less than 0.5 false alarms per day .
•Here, we monitor apatient for 1year,and report the results.
Materials & Methods
•A patient (age 15y) with Dravet Syndrome wore the Embrace smartwatch, which sent CS alerts to a smartphone for generating calls to
caregivers, and logged the calls and alerts time in an online database. The patient was never left alone and the patient’s caregivers were
asked to meticulously report each CS.
•Two data experts independently inspected all the data to further verify the accuracy of caregivers’ self-reports.
•Sensitivity was computed as the percentage of CSs that automatically triggered an alert. False alarms were counted by subtracting the
number of correctly recognized CSs from the total alerts.
•For the first time, we have
tested the Embrace seizure
detection system during 1year
in real-life settings.
•The performance is in the same
range as previous tests in best-
case clinical settings, i.e.
Sensitivity>95%and false alarm
rate<1alarm per day .
•Similar longitudinal analysis on a
more heterogeneous patients’
population is underway.
Alert generation and
Alert delivery to
Figure 1.Top:Schematics of Embrace CS detection and alert system, composed
of awristband that detects an event, transmits an alert to his/her smartphone,
which generates acall via acloud-based service to designated caregivers.
Bottom:Detailed schematic workflow of the CS detector wristband.
 Poh et al 2012, Epilepsia, 53(5),93-7
 Regalia et al 2015, American Epilepsy
 Onorati et al 2016, Epilepsy Pipeline
 Onorati et al 2016, European Congress
 Picard et al 2016, American Epilepsy
 Caborni et al 2016, Partners against
Mortality in Epilepsy conference
Figure 3.(A) Overview of Embrace outcome during the
monitoring period (one row = 1 day).Green:timing of the 45
detected seizures;purple:timing of the missed seizures;red:
timing of false alarms;dark blue:detected rest periods;light
blue:estimated rest period;gray:device off-wrist.
(B) Distribution of seizures (top) and false alarms (bottom)
wrt clock time.
•Over aperiod of 11.3 months (341 days),
the patient wore Embrace 5,968 hours
(17.5±3.5 hours per day) and experienced
•Of the 46 CS’s, 45 produced real-time alerts
(Sensitivity=98%)(examples in Figure 2;
green ticks in Figure 3A).
•All true positives were successfully
transmitted to the phone (mean delay=13
sec) and to the online database (mean
•The overall false alarm rate was 0.14 per
day worn (51 false alarms over the year;
red ticks in Figure 3A).
•Zero false alarms were triggered during
rest periods (blue segments in Figure 3A).
•Most days had zero false alarms (315 out of
341 days, i.e. 92%)
Figure 2.Example of electrodermal activity (EDA) and
acceleration (ACC) recorded by Embrace during two
convulsive seizures (left:detected;right:missed).
Orange bar refers to the detection time (left) or to the
seizure time reported by the patient (right).