Information Technology Reference
In-Depth Information
20.4.1 Algorithm for Generating Automatic Warnings about
Impending Seizure from EEG
Based on the methodology, presented in the previous section, the algorithm for gen-
erating automatic warnings about possible seizure from multichannel EEG record-
ing can be outlined as follows. The algorithm consists of two key phases, namely the
training and the seizure detection. During the first stage, the EEG data are recorded
and analyzed to determine the critical sites with respect to the brain's transition
into seizure, which are individual to each patient. The sites found during the train-
ing phase are used to automatically detect the conditions signalizing of impending
seizures. The algorithm for generating automatic warnings includes the following
steps:
•
Training phase:
1. Collect EEG data for a given number
m
≥
1 of the first seizures detected with
manual assistance of a qualified person;
2. For each seizure, determine a given number
k
of critical electrode sites by
following steps:
-
estimate the STLmax values for all electrodes in a 10-min window imme-
diately before the seizure
T
ij
-
compute matrix
Q
=(
)
i
,
j
of T-indices of the STLmax values between a
of electrode sites
- solve the corresponding quadratic 0-1 problem (20.7)
3. Among
m
different
C
i
,1
pair
(
i
,
j
)
m
sets of critical sites, select:
- either the sites that are common to all
m
seizures, i.e.,
1
≤
i
≤
m
C
i
- or electrode sites that can be found in most seizures
≤
i
≤
•
Seizure alert phase:
1. Compute the critical threshold value
t
α
/
2
,
d
−
1
, where
d
is the total number of
the STLmax values per channel in 10-min window
2. Sequentially analyze the EEG from the electrode sites selected in Step 3 of
the training phase as follows:
-
calculate STLmax values
compute corresponding T-indices
T
ij
-
(
t
)
between pairs of selected critical
sites
go to the next step (Step 3 below) when the T-indices
T
ij
-
(
t
)
drops below
the threshold
t
α
/
2
,
d
−
1
, i.e.,
T
ij
t
α
/
2
,
d
−
1
- otherwise continue sequentially analyzing the data
3. If the threshold-drop time
t
lies within some fixed prediction horizon
h
of the
previous warning, then go back to the previous step (Step 2 of the seizure alert
phase); otherwise generate a warning.
The proposed algorithm is a version of the adaptive threshold seizure warning
algorithm (ATSWA) introduced by Sackellares et al. in [35]. In particular, the main
difference between the new version and ATSWA is that the proposed version utilizes
(
t
)
>
