Wrist accelerometry-based detection of nocturnal myoclonic seizures

Full text

Wrist accelerometry-based detection of nocturnal myoclonic seizures
F. Onorati1,L.Babilliot1,G.Regalia1,R.El Altrache2,R.Picard1,3
1. Empatica,Inc, Boston, MA, US; 2. Boston Children’s Hospital, Boston, MA, US; 3. Massachusetts Institute of Technology, Cambridge, MA, US
Background
Materials and Methods
Results
Both
ambulatory
and
outpatient
applications could benefit greatly from robust and reliable non-EEG
wearable devices
that automatically
detect
and
monitor
for
epileptic seizures
.
To date, the only FDA-cleared wrist-worn seizure monitor focuses on detecting only generalized tonic-
clonic seizures (GTCSs)
[1]
.However, the
accelerometer
(ACM) sensor embedded in
wearable devices
provides an
inexpensive
,
comfortable
and
easy-to-use
solution for sensing and characterizing other
types of motor movements
[2]
.
Myoclonic seizures
(MS) are focal or generalized motor seizures characterized by
short sudden
shock-like movements
. They occur typically during
sleep
or
after periods of rest
. Clusters of MS can
precede more severe seizures
like GTCSs. Myoclonic epilepsies can be
refractory
to drug treatment.
State-of-the-art
algorithms based on
ACM
for
MS detection
report
sensitivity
(Sens) as high as
0.8
but at acost of
false alarm rates
(FAR) as high as
100 per hour
,even during
sleep [3]
.
•In order to compute the
Specificity
,the
Precision
and the
F1-score
,the events identified at the segmentation phase
(see
Figure 2
) have been considered as positive (MS, NMS =34) and negative (non-MS, NNMS = 1581) events.
•At the
optimal point
, computed in order to maximize the
F1-score
, the algorithm detected
23 MS
, reaching
Sens = 0.735
(red square in
Figure 3
).
Data
1. TS1(inpatient)
Collected with
Empatica E4
in an Epilepsy
Monitoring Unit (
EMU
)at Boston Children’s
Hospital (Boston, MA) where
MS
were
annotated from v-EEG examination by
a
board certified clinical neurophysiologist.
#Sbj (w/ MS) #MS Total
hours
TS1
4 (4) 34 432
TS2
4 (0) 0108
Table 1.
Summary of the datasets
Figure 3.
ROC curve analysis of the MS detection algorithm.
Sens and FAR at different values of the decision threshold on
the algorithm probability output. Conveniently, The
performances at Sensitivity = 0.7, 0.8 and 0.9 have been
highlighted. The red squared point refers to the optimal
threshold found maximizing the F1-score.
Figure 1.
Examples of cluster of myoclonic seizures (left) and non seizure event (right), identified after the segmentation
phase (see Figure 2). The violet area indicates the label as identified by a neurologist through v-EEG inspection. Red
dashed lines delimit the events as segmented by the algorithm. The black line shows the pattern recognition output
performed during the data reduction phase (see Figure 2).
2. TS2 (outpatient controls)
From healthy subjects collected by
Empatica Embrace
during
sleep
as identified by a
proprietary algorithm
.
Algorithm
A
proprietary MS detection algorithm
was used to classify sensor data into MS or non-
MS events.
Analysis
A
leave- N-seizures-out cross-validation
(N = 3) approach was employed in order to
compute the performance of the algorithm.
Performance was measured in terms of Sens, FAR, Specificity, Precision and F1-score.
Two
datasets
were used for algorithm development and evaluation:
Overall MeanCV MinCV MaxCV
F1
-score
0.633 0.657 0.400 0.857
Sensitivity
0.735 0.767 0.335 1.00
Specificity
0.987 0.989 0.965 1.00
FAR
1.13/night 0.96/night 1.46/night 0.6/night
Precision
0.556 0.672 0.444 1.00
Table 2.
Summary of the performance
•The performance computed with the cross-
validation approach proved to be robust as the
minCV and maxCV values of the single performance
metrics do not vary much with respect to the
meanCV and overall values.
Conclusions
Promising preliminary results
have been shown with the development of a new
automated nocturnal MS detection
algorithm
using
machine learning
and
ACM data
recorded from a wrist-worn wearable device (Empatica E4 or Embrace).
The
Sensitivity
is close to the
current state-of-art
presented in the Literature, while the
FAR
shows a significant
reduction, reaching values more
bearable
for users and caregivers.
The data collection approach (i.e., two different dataset for MS and non-MS events) was intended to overcome the
difficulty
for the labeler to provide a thoroughly
ground truth
, resulting in the presence of possible
unlabeled MS
in the
inpatient dataset. A
v-EEG revision
is ongoing for future development.
[1]
Onorati 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]
Ulate-Campos et al. (2016). Automated seizure detection systems and their effectiveness for each
type of seizure. Seizure, 40,88–101.
[3]
Nijsen et al. (2010). Time-Frequency Analysis of Accelerometry Data for Detection of Myoclonic
Seizures. IEEE Transactions on Information Technology in Biomedicine, 14(5), 1197–1203.
References
Contacts Francesco Onorati
fo@empatica.com
Rosalind Picard
rp@empatica.com
Segmentation
Data
Reduction Processing Feature
Extraction Classification
ACM MS
Non-MS
Figure 2.
Pipeline of the different phases of the algorithm. During the “Data Reduction” phase, a
pattern matching algorithm is performed in order to identify possible events of interest while
dramatically reducing the data to process in the successive phases.
AIM: Present preliminary results of a new automated machine-learning algorithm to
detect nocturnal myoclonic seizures (MS).