IEEE Computational Intelligence Magazine - November 2020 - 19

and 64 × 64 from 150 CT images. To increase the classification
performance, the feature extraction process was performed on
each patch. Five computational learning theory algorithms were
adopted and utilized as feature extraction methods: a gray level
co-occurrence matrix (GLCM), a gray level run length matrix
(GLRLM), a local directional pattern (LDP), a discrete wavelet
transform (DWT), and a gray-level size zone matrix (GLSZM).
To avoid an overfitting problem, k-fold cross validation was performed during training. With GLSZM and 10-fold cross-validation, the classifier achieved the best accuracy (99.68%).
Randhawa et al. [47] proposed a method of using computational learning theory for genome analyses. This method combines decision trees with digital signal processing to construct a
model for classification of the COVID-19 virus sequences and
can identify intrinsic viral genomic signatures. To validate the
results of identifications, Spearman's rank correlation coefficient
analysis was adopted. The proposed method can be used to analyze large datasets containing more than 5,000 unique viral
genomic sequences. In this dataset, there are 29 COVID-19
viral sequences, implying an imbalanced data issue (29: 5000).
The proposed method achieved a 100% accuracy. Furthermore,
the proposed method uses only raw DNA sequence data to discover the most relevant relationships between more than 5,000
viral genomes within minutes from scratch. This shows that, for
new viral and pathogen genome sequences, unmatched
genome-wide machine learning methods can provide reliable
real-time courses of action for taxonomic classification.
Mei et al. [48] developed an ensemble model to identify
COVID-19 infections, which can allow early identification of
COVID-19 patients at an early stage based on the initial chest
CT scans and related clinical information. This model combines a deep convolutional neural network with three classifiers: random forest, support vector machine, and multilayer
perceptron. The deep convolutional neural network is utilized
for imaging the characteristics of COVID-19 patients, and the
three classifiers form an ensemble model to classify COVID-19
patients based on extracted characteristics of COVID-19 and
other clinical information. This ensemble model showed significant performance in terms of sensitivity (84.3%), specificity
(82.8%), and AUC (0.92).
Apostolopoulos et al. [49] extended their previous work
[19] by using transfer learning to train deep CNNs since
there are many pre-trained models that can be retrieved from
open sources, such as VGG-19 [52], MobileNets V2 [30],
Inception V4 [53], and Xception [54]. Unlike their previous
work [19], which straightforwardly utilized MobileNets V2
to build an image recognition model for classifying COVID19 patients, Apostolopoulos et al. [49] applied transfer learning on the pre-trained models and used a dataset that consists
of 224 chest CT images of patients with COVID-19, 700
chest CT images of confirmed common bacterial pneumonia, and 504 chest CT images of no diseases to fine-tune the
pre-trained models.
Based on the papers surveyed in this section, some insightful
findings can be made:

1)	Although deep learning has become the most popular
notion recently, some classical computational learning
theory approaches, such as support vector machine, random forest, and decision tree, could still be useful while
the amount of data is limited. The studies [46]-[48] reveal
that the shallow learning method can be utilized as an
initial model for building a classification model to distinguish COVID-19 patients.
2)	 With a bigger dataset, the concept of model ensembles can
be used to combine initial models with some deep learning
methods. The work in [45] and [47] provides possible solutions for model ensembles. Besides model ensemble, the
concept of domain adaptation is a possible solution to combine two models. The transfer learning techniques utilized in
Apostolopoulos et al.'s study [48] are also a possible solution.
VII. Probabilistic Methods for Combating COVID-19

In computational intelligence, a probabilistic method is applied
by calculating the expected value of a random variable. The
probabilistic method is typically used for analysis of the risk
factors correlated with COVID‐19 and explains why they are
crucial. Table VI shows the issues that have been addressed by
probabilistic methods.
Cássaro and Pires [55] assume that the number of infected
patients grows exponentially over the time. As a result, the probabilistic model can be formulated as
I ^ t h = I ^t 0h e rt, (3)

	

TABLE V Issues addressed by existing computational
learning theory methods.
TPVP
DUFFEY AND ZIO'S
STUDY [44]

CSVI

TRD

PD

PHPM



WANG ET AL.'S
STUDY [45]



BARSTUGAN ET AL.'S
STUDY [46]



RANDHAWA ET AL.'S
STUDY [47]



MEI ET AL.'S
STUDY [48]



APOSTOLOPOULOS
ET AL.'S STUDY [49]





TABLE VI Issues addressed by existing probabilistic
methods.
TPVP
CÁSSARO AND
PIRES'S STUDY [55]



ZHANG ET AL.'S
STUDY [57]



KUCHARSKI ET AL.'S
STUDY [56]



CSVI

TRD

PD

PHPM

NOVEMBER 2020 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

19
IEEE Computational Intelligence Magazine - November 2020

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - November 2020
