Content uploaded by Francesco Onorati
Author content
All content in this area was uploaded by Francesco Onorati on Feb 13, 2020
Content may be subject to copyright.
Prospective evaluation of generalized tonic-clonic seizure multimodal detection:
comparison between pediatric and adult cohorts in EMU
F. Onoratia, G. Regaliaa, W. Curt LaFrance Jrb, A. S. Blumc, J. Bidwelld, P. De Lisoe, R. El Atrachef, T. Loddenkemperf, F. Mohammadpour-Touserkanif, R. Sarkisg, D. Friedmanh, J. Jeschkeh, R. Picarda,i
•The
“fixed and frozen” AI-based GTCS detection algorithm
complies with
FDA requirements
(lower bound of CI for PPA>70% and
FAR<2) for both pediatric and adult populations.
•The
FAR for pediatric
is significantly (p-value<0.01)
higher FAR
,most likely because children were
more active
in the EMU.
•
During rest
, the overall
FAR drops dramatically
to
0.03 FA/24h
, which is as low as
0.01 FA/night
(considering 8 hours of sleep per
day), i.e.
1FA every 100 nights of sleep
.
•
Future work
will examine the
prospective detection performance
and overall impact in
outpatient settings
.
Results
Conclusions
[1]
Onorati, F., Regalia, G., et al. (2017). Multicenter clinical assessment of improved wearable multimodal convulsive seizure detectors. Epilepsia,
58(11), 1870–1879.doi:https://doi.org/10.1111/epi.13899
[2]
Regalia et al. (2019). Multimodal wrist-worn devices for seizure detection and advancing research : Focus on the Empatica wristbands.
Epilepsy Res. 153:79 82.doi:10.1016/j.eplepsyres.2019.02.007.
[3]
510(k) Premarket Notification (K181861), https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfPMN/pmn.cfm?ID=K181861,12/20/2018.
[4]
Regalia, G., et al. (2019). Sleep assessment by means of awrist actigraphy-based algorithm: agreement with polysomnography in an
ambulatory study on older adults,
in preparation
.
References
Table 1
and
Figure 2
show the performance as measured in the
prospective study
presented for the
FDA clearance
.
a. Empatica Inc, Boston, MA, USA
b. Division of Neuropsychiatry and Behavioral Neurology, Rhode Island Hospital, Brown University, Providence, Rhode Island, USA
c. Department of Neurology, Rhode Island Hospital, Brown University, Providence, Rhode Island, USA
d. Harvard Medical School, Boston, MA, USA
e. Neurology Unit, Department of Neuroscience and Neurorehabilitation, Bambino Gesù Children’s Hospital, IRCCS, Rome, Italy
Contacts
Francesco Onorati
fo@empatica.com
Rosalind Picard
rp@empatica.com
f. Department of Neurology, Boston Children’s Hospital, Boston, MA, USA
g. Department of Neurology, Brigham and Women's Hospital, Boston, MA, USA
h. Department of Neurology, New York University Langone Medical Center, New York, New York, USA
i. MIT Media Lab, Massachusetts Institute of Technology, Cambridge, MA, USA
•
Embrace
wristband (Empatica Inc., Boston, MA) runs an
AI-based GTCSs detection algorithm
that can trigger an alert to designated
caregivers for prompt intervention (Figure 1).
•The combination of wrist
accelerometer
(ACM) and
electrodermal activity
(EDA) has been shown effective in the
detection of
generalized tonic-clonic seizures
(GTCSs), with sensitivity > 94 % and 1 false alarm every 5days on retrospective studies
[1][2]
.
Rationale
Figure 1 Embrace wristband detects an event and transmits an alert to a smartphone, which generates notifications via a cloud-based service to designated caregivers.
•To assess the
effectiveness and safety
of the seizure detection device,
FDA
required a
prospective study
where a predefined
detection algorithm (i.e.,
“fixed and frozen”
)is applied to an
EMU test cohort [3]
.
#GTCS
(detected)
PPA (cPPA) [95% CI]
Detection
Delay [s]
#FA (@ Rest)
#Days (@ Rest)
FAR (@ Rest)
All
54 (53) 98.1% (94.8%) [90.3% -97%] 38.19 383 (6)
408.56 (200.93)
0.94 (0.03)
6
-
21
32 (31)
96.9% (91.5%) [83.4%
-
95.3%]
36.9 218 (4) 161 (82.1) 1.35 (0.049)
>21
22 (22) 100% (92.4%) [91% -93.1%] 40 165 (2)
247.56 (118.83)
0.67 (0.017)
Figure 2 Histograms of FA, days of recordings and FAR across all the patients in the study.
Table 1 Prospective performance analysis of the on the test cohort
We present the results of a prospective inpatient analysis of Embrace wristband. FDA designated
endpoints are met for ages 6 and up. The average number of false alarms is close to zero per night.
Endpoints
•
Primary endpoint
: the lower bound of the
95%confidence interval (CI) of positive percentage agreement (PPA)
is higher than
70%
.
•
Secondary endpoint
: the
false alarm rate per 24h
(FAR) is lower than
2FA/day
.
Materials & Methods
Secondary endpoint -FAR
•The
overall FAR
was found to be
0.94 FA/24h
;
Pediatrics
FAR was
1.35 FA/24h
;
Adults
FAR was
0.67 FA/24h
.
•The percentage of
patients with 0FAs
was
44%
for
pediatrics
and
45%for adults
.
•The
median FAR
was
0.2 FA/24h
(0.24 for
pediatrics and 0.17 for adults).
•
41 FAs
(18 from pediatrics) were
seizures
just not
of type GTCS
.
Omitting these, overall corrected
FAR (
cFAR
)decreases to
0.84 FA/24h
.
•During
rest periods
, the
FAR
dropped to
0.03
FA/24h
(0.049 for pediatrics and 0.017 for adults).
Primary endpoint -PPA
•Overall,
53 out of 54 GTCSs
were detected
(PPA =98.1%, cPPA =94.8%).
•For
pediatrics
,
31 out of 32 GTCSs
were
detected (PPA =96.9%, cPPA =91.5%)
•For
adults
,
22 out of 22 GTCSs
were detected
(PPA =100% and cPPA =92.4%).
Data
•
141
patients diagnosed with epilepsy were studied in the
EMU
at
6clinical sites
using v-EEG:
80 pediatrics
(36 females, 6-21 years)
and
61 adults
(33 females, 22-63 years).
•Overall,
409 days
(9,806 hours, 5,727 from pediatrics) of data were collected (median:49.23 hours/patient);
•
Three board-certified neurologists independently
marked seizures from
v-EEG
.
Majority rule
(i.e., at least 2 out of 3) was used for
the final labels of GTCS or non-GTCS.
•
31 patients
(17 pediatrics) experienced a total of
54 primary or secondary GTCSs
(17 pediatrics), while 110 patients experienced no
GTCS.
194 non-GTCS
also occurred.
•
Rest
accounted for
51%
of
pediatrics’
and
48%
of
adults'
data.
Analysis
•The
“fixed and frozen” GTCS detection algorithm
was used to prospectively analyze the data.
•PPA is also reported as corrected PPA (
cPPA
), due to the presence of
multiple seizures per patient
.
•The
PPA 95%confidence limits (CI)
were computed with a
non-parametric bootstrapping approach
.
•The
FAR
was evaluated on
all the recordings
and by considering only the periods of
rest
as identified by a
proprietary algorithm [4]
.