IEEE Systems, Man and Cybernetics Magazine - January 2022 - 37

for the linear support vector to find planes for segmentation.
This is why we see an increase in accuracy, precision, recall,
and the F1 score. In contrast, we see accuracy decreasing for
the RBF model with SMOTE ENN because of the linearity
introduced among the minority class by SMOTE ENN, which
makes RBF challenging to classify.
Tree-Based Algorithms
The accuracy of the decision tree was significantly high
compared to other models, even with the unbalanced
data set. SMOTE ENN increased its accuracy slightly
because of thee linearity introduced by SMOTE ENN.
Precision, recall, and the F1 score resulted in the same
trend as accuracy when the SMOTE ENN was applied for
the same reason
As AdaBoost is just an ensemble method of the decision
tree, it shows the same trend as the decision tree. The
results of the AdaBoost are slightly better than the decision
tree because it is an ensemble of multiple decision trees.
MCNNs
SMOTE ENN increases the ability for moving averages to
find smoother representations of data, and thus its accuracy
increases slightly. As the SMOTE-ENN algorithm created
linearity in positive examples, it becomes easier for the
MCNN to pick true positives, thus explaining the increase
in precision. Compared to the other ML algorithms, the
MCNN achieved a lower performance score due to the limited
number of training examples of 1,990 and 3,656 for the
unbalanced and balanced data sets, respectively.
Conclusion
The work presented in this article investigated comparisons
of various algorithms and selected the best algorithm
for each data set. Table 7 illustrates the best performance
in each category. We see that AdaBoost performed the best
in the majority of the categories. AdaBoost is the best algorithm
in terms of accuracy for both the original unbalanced
and balanced data sets using SMOTE ENN.
Naïve Bayes' performance is the best in terms of precision
with the original data set. In the case of the balanced
Table 7. The best performance for
accuracy, precision, recall, and F1 score
for the original and balanced original
data sets, respectively, using the SMOTEENN
algorithm.
Best
Performance Accuracy Precision
AdaBoost
Original
data set
Original
data set
with
SMOTE-ENN
AdaBoost
Naïve
Bayes
Nearest
neighbors,
QDA
Recall
AdaBoost,
linear SVM
F1 Score
AdaBoost
AdaBoost Gaussian
process
data set, nearest neighbors and QDA achieved the highest
precision. On the original data set, AdaBoost and
the linear SVM both realized the best recall, even
though AdaBoost maintained the best recall for the balanced
data set.
Again, AdaBoost generated the best F1 score compared
to the other ML algorithms on the unbalanced data set.
However, the Gaussian process achieved a higher F1 score
on the balanced data set. From this study, it can be concluded
that AdaBoost is the most invariant to distribution
among the examined algorithms.
Main Contribution
The main objective of this study is the investigation of the
utility of different ML algorithms to classify hypertension
episodes. The following are the main contributions of
this article.
◆ Nine ML algorithms with minimal computational
requirements were developed to analyze ICP signals
and predict hypertension episodes.
◆ An analysis was conducted of the performance and
robustness of the SMOTE-ENN algorithm for predicting
hypertension signals in an unbalanced ICP data set
with limited labeled data points.
The Direction of Future Work
The future direction of this work is a further investigation
of many different forms that represent the same data set
and explore effects in the classification ability of various
classifying algorithms. The possible different representations
can be cylindrical- and spherical-coordinate systems,
homogeneous coordinate system, curvilinear coordinates,
orthogonal coordinates, skew coordinates, log-polar coordinate
system, and many others.
The intuition behind exploring the different representations
of the data set for various algorithms is that the algorithms
may match easily with the data set representation.
This could result in better classifying metrics.
Additionally, more advanced data-driven ML algorithms
can be explored with the procurement of more
labeled data. Higher-capacity algorithms can provide better
state-of-the-art performance but require significantly
more data to be used for training. Additional ICP data can
be gathered and labeled so that more complex algorithms
can be explored.
About the Authors
Arif Jahangir (ajahangir@ryerson.ca) is with the Department
of Computer Science, Ryerson University, Toronto,
Ontario, M5B 2K3, Canada.
Kavyan Tirdad (tirdad@ryerson.ca) is with the Department
of Computer Science, Ryerson University, Toronto,
Ontario, M5B 2K3, Canada.
Alex Dela Cruz (adelacru@ryerson.ca) is with the
Department of Computer Science, Ryerson University,
Toronto, Ontario, M5B 2K3, Canada.
January 2022 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE 37
IEEE Systems, Man and Cybernetics Magazine - January 2022

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