IEEE Systems, Man and Cybernetics Magazine - January 2022 - 32

A monitoring device was installed in the Robert Wood
Johnson Medical Center of Rutgers University's surgical
intensive care units. Data-gathering units were coupled
with patient monitors. The device was activated as soon as
a patient arrived with a diagnosis of brain injury that
requires an ICP bolt/ventriculostomy. Individuals were
mostly ventilated and sedated. ICP was uninterruptedly
monitored with ventriculostomy.
A General Electric TRAM-RAC 4 A was used to acquire
the outputs from clinical monitors. The rate of sampling was
50 Hz, with a resolution of 1.41 mV at ±5 V, which corresponds
to the pressure resolution of 0.14-mm Hg and a range of
±500-mm Hg. ICP was uninterruptedly recorded using microtransducers
(a Camino direct pressure monitor, Camino Laboratories,
San Diego, California, USA), which were implanted
intraparenchymal into the frontal cranium.
Data Preprocessing
The data samples were first analyzed to identify negative
measurement readings of higher than 100 Hg. Negative values
or values higher than 100 Hg were caused by the movement
of the monitoring device and do not provide
information about ICP. Negative values or values higher
than 100 were encoded with a not-a-number (NaN) value.
NaN values were subsequently replaced by interpolating
the values 50 steps before and 50 steps after the NaN
value. Fifty steps prior and after were chosen for obtaining
a smooth spline for interpolation. From the data set, sampling
was done at 50 Hz.
Identification of Hypertension Signals
Hypertension signals were identified by scanning for values
greater than 20 Hg. Six minutes of data before the designated
point were identified, extracted, and labeled as
positive samples. This was done for both the MIMIC II and
CHARIS data sets. The remaining data signals were divided
into 6-min durations and labeled as negative samples.
Six minutes was deemed suitable due to the following considerations.
◆
The length of the ICP signals collected from the MIMIC
II and CHARIS data sets should be sufficient for both
the training and testing phases.
◆ The clinical response time that is required for medical
staff to take necessary actions to avoid the HT state of
patients.
Positive and negative samples were then combined into
one data set for both the MIMIC II and CHARIS data sets.
Data Balancing
The data set contained unproportionally more negative
samples (no hypertension signals) compared to positive
(hypertension episode). The data set was then treated with
an SMOTE-ENN algorithm to balance the data. SMOTE
oversamples the data set by creating a synthetic data point
for the minority class. It takes a vector from one of the
minority-class k neighbors. It then multiplies a random
32 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE January 2022
number x, which is between one and zero. This is complemented
to the current data point to create a new artificial
data point.
For individual point p in the rear class,
1) calculate its k-nearest neighbors (KNNs) in the rear
class
2) arbitrarily choose rk# of the neighbors, which is done
by replacement
3) pick a random point p along the line and then join p and
each of the r selected neighbors
4) complement these fictitious points to the data set with
the rear class.
After the SMOTE algorithm, data was passed through
the ENN method. ENN creates a forecast in two-way communication.
It takes into account the nearest neighbors of
the test sample and considers the test sample as their nearest
neighbors. It exploits the universal classwise statistics
from all training data by iteratively assuming every probable
membership class of a test sample. Given an unknown
sample X to be classified, it is iteratively assigned to each
class
jN
,, ,
according to (1):
C
ENNV112,,,, 1
!
= argmaxjN(|c
f
ij
N
D +=
()i +
nk
nk kT
j
i
iim
-|c
ij
N
!
D
j
i
nk
n
i
m1,
(1)
where k is the user-defined parameter for nearest neighbors;
ni is the number of training data for class
ik , i =
number of nearest neighbors of test sample X for class
iT , i = generalized classwise statistics of class I; and i
D
j
is the change of KNNs for class i when test sample X is
labeled as belonging to class j.
ML Approaches
Traditional ML Algorithms
Traditional ML algorithms are primarily based on statistical
theorems, providing a simplistic implementation with
the limited number of learning parameters a model needs
to learn during training. This allows traditional ML algorithms
to be trained on a limited sample size (hundreds)
compared to more complex algorithms with a large number
of parameters (millions), such as state-of-the-art neural
networks, which require a large volume of data to be
trained (generally tens of thousands or more).
Nearest Neighbors
These types of algorithms are nonparametric classifiers. A
KNN algorithm is regarded as the simplest classification
method in ML and can be considered as lazy learning or an
instance-based classification method. It classifies using a
voting strategy based on the KNN using a distance metric
d (typically Euclidean distance). A weight voting strategy
is applied that weights the vote of each neighbor based on
their observed distance / .d1
=12 f and predicts the class membership
IEEE Systems, Man and Cybernetics Magazine - January 2022

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - January 2022
