Conference PaperPDF Available

Real-time Performance in Outpatient Settings for Embrace Seizure Detection System with User-Adjustable Sensitivity

Authors:

Abstract

Clinically validated automated mobile seizure detection devices have been poorly tested in outpatient settings to date. Detectors relying on motion-related signals to recognize convulsive seizures (CSs) are likely to yield higher false alarm rates (FAR) in real life than in a clinical environment, due to a wider range of possible physical activities. Embrace is a wrist-worn device combining Accelerometer and Electrodermal Activity to detect and alert to CSs. Retrospective analysis combining machine learning classifiers trained on outpatient data (Onorati et al, 12th European Congress on Epileptology, 2016), and different sensitivity levels for active and rest periods (Caborni et al, 32nd International Epilepsy Congress, 2017), yielded a FAR of less than 1 false alarm (FA) per day. The aim of this work is to investigate Embrace real-time performances when allowing user-adjustable sensitivity level in outpatient settings.
Embrace
(Empatica Inc., Cambridge, MA) is a
wrist-worn
device combining 3-axis
accelerometer and electrodermal activity sensors
to detect and alert
for
epileptic convulsive seizures
(
Figure 1
).
Real-time Performance in Outpatient Settings for Embrace Seizure Detection System
with User-Adjustable Sensitivity
G. Regalia1,C.Caborni1, R.W. Picard1,2, F. Onorati1
For the
first time
,we evaluated if the
user behavior could help improve
Embrace system performances
.
The
user-adjustable Sensitivity mode
provided an overall benefit in
improving nocturnal seizure detection
(better than ‘always Sens1’)
while
keeping bearable FARs
(better than ‘always Sens2’).
Not all users complied with the suggested guidelines, which reflects
different individual needs and typical seizure times
.
Additional analysis are underway evaluating a
larger pool of users
for
l
onger monitoring periods
and
relating performances to the activity level
of
each subject.
1. 10 users
were involved in this testing. They were asked to
switch the Sensitivity levels
using the Alert App according to
these criteria:
High activity time
àSens1(i.e., standard mode).This configuration exhibits
reduced False Alarms
at alower level
of overall Sensitivity.
Rest/low activity time
àSens2.This configuration exhibits
increases probability of detecting seizures during
sleep
,when the chance of false alarms is likely lower.
2. The users and their caregiver were asked to
meticulously report all the experienced seizures
.
3. The dataset consists of
1,288 hrs
of
outpatient
recordings (median hrs/patient=144,min=36,max=187).
4. Sens (# recognized seizures/# total seizures) and FAR (# False Alarms/24 hrs or #False Alarms/1hr)were computed
on all the recording period and with respect to rest and active times.
The
user compliance
was fairly good:
5 out of 9
users switched to ‘Sens2during
rest time
, as suggested;
5 out of 10
users switched to 'Sens1' for
most of their
active time
(
Figure 5
).
Three patients
experienced
4 convulsive seizures (CSs)
, including 3 nocturnal seizures (red flashes in
Figure 5
).
The
overall real-time
performances were
Sens=100%
(4/4 CSs) and
FAR=0.48/24 hrs
(26 FAs in total) (red diamond in
Figure 3
).
No FAs
were
triggered during rest (
Figure 5
, bottom).
Keeping fixed Sensitivity levels
(off-line simulations) showed
worse performances
(
Figure 3
): using always Sens1 resulted in 2 missed nocturnal
CSs, while using always Sens2 doubled the FAR.
1 user (user 4) who kept Sens2 also during high activity time would have her
individual FAR decreased by >70% using
Sens1 (
Figure 4
).
1. Empatica Inc,Milan, Italy
2. MIT Media Lab, Massachusetts Institute of Technology, Cambridge, Massachusetts, U.S.A
Contacts
Francesco Onorati
fo@empatica.com
In this work we explored how the performance of Embrace in outpatient settings change,
when the users are allowed to adjust the Sensitivity level on the fly.
Aim
Figure 1
Embrace convulsive seizure detection and alert system, composed of a wristband that detects an event, transmits an alert to a smartphone, which generates a call via a cloud-
based service to designated caregivers.
40
60
80
100
0,0 0,2 0,4 0,6 0,8 1,0 1,2
Sensitivity (%)
FAR (per 24 hrs)
Sens1
for all
users (off-line)
Sens2
for all
users (off-line)
Real-time
performances
Rosalind Picard
rp@empatica.com
Chiara Caborni
cc@empatica.com
Figure 3
Cumulative performances across users in terms of FAR and
Sensitivity, computed on all the recorded data and compared
to off-line simulations with constant Sensitivity level.
[1]
Onorati et al., 2016,Improving convulsive seizure detection by exploiting data from outpatient
settings using the Embrace wristband, 12th European Congress on Epileptology, at Prague.
[2]
Caborni et al., 2017,Convulsive Seizure detection improved by setting different parameters for
active/rest periods, 32nd International Epilepsy Congress, at Barcelona.
[3]
Onorati, Regalia et al., 2017, Multi-center clinical assessment of improved wearable multimodal
convulsive seizure detectors, Epilepsia, 8(11):1870-1879.
References
Figure 2
Accelerometer signals, activity segmentation and user-adjusted Sensitivity levels. Rest segments were detected with a proprietary
algorithm, while high activity segments were defined as when standard deviation of accelerometer magnitude exceeded 0.1 g.
4
38
2
2
3
Figure 5
Usage of ‘Sens1’ or ‘Sens2’ level during active
(top panel) and rest (bottom panel) time for each
user. Detected seizures are in red, False Alarms
are in black.
Figure 4
Individual FAR obtained during active time for each user (red), compared to off-line
simulations with a constant Sens level .
Detected seizure False Alarm
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
1 2 3 4 5 6 7 8 9 10
FAR (per hr)
user ID
FAR during active time
user-adjusted
'always Sens 1'
'always Sens2'
Results
Compared to clinical environments, wearable detectors relying on motion-related signals are likely to yield
more false alarms in outpatients settings
,
due to the higher frequency of
“seizure-like” physical activities.
We have previously tackled this challenge by
training machine learning classifiers on both clinical and outpatient data[1]
and by simulating
different
Sensitivity levels for active and rest periods[2]
.We obtained a
False Alarm Rate (FAR) lower than 1/24 hrs,
while keeping
Sensitivity (Sens) high
(
>90%
), thus approaching performances of similar detectors in clinical setting validation
[3]
.
Rationale Methods
Conclusions
Giulia Regalia
gr@empatica.com
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.