Добавил:
Sekretar
kiopkiopkiop18@yandex.ru
t.me/Prokururor I Вовсе не секретарь, но почту проверяю
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз:
Предмет:
Файл:Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_145_библиотеки_им_акад_М_И_Перельмана
.pdf
Detection of Disease Severity Disease Prediction using Machine Learning 33
https://t.me/med1917
collection mechanisms are inconvenient to patients, time-consuming, and require
high cost. Thus, there is a need to integrate sensor-based data collection,
transmission to cloud storage, and implementation of ML models on the data
stored on cloud storage for disease diagnosis and severity detection.
Table 1. Comparative analysis of existing DL-based techniques used for disease detection and severity
prediction.
Authors, Year,
and Citation
Yao et al., 2020
Li Yan et al.,
2020
Tadesse et al.,
2020
Tripoliti et al.,
2017
Liu et al. KNN, and SVM, PhysioNet dataset
Technique Applied
Predictive Logistic
Regression Support
Vector Machine,
Random Forest, KNearest Neighbour
(KNN), and
AdaBoost
Multi-tree XGBoost
model for
categorizing critical
and severe cases of
COVID-19
SVM
Review of KNN,
SVM, for prediction
of severity of heart
disease and heart
rate failure
Dataset used
Health parameters
recorded by urine and
blood test of patients
of COVID-19, and
their clinical features
such as age, blood
pressure, heart rate
etc.
Selected three
features viz. Lactic
Dehydrogenase
Lymphocyte, and
high sensitivity C-
reaction protein from
375 patients of
COVID-19
ECG signals from
patients of Hand Foot
and Mouth
Disease(HFMD) and
tetanus.
PhysioNet dataset
Performance
Severity
prediction
accuracy of
99.17% on the
training dataset,
and 81.48% on
the testing dataset.
Predicted the
survival rate with
the accuracy of
approximately
90%
98.1% accuracy
for HFMD, and
78% for the
tetnus.
Best accuracy of
96.39% by using
KNN
Claimed accuracy
of 100% for KNN
Challenges
Conclusions made using
a small dataset confined
to only one geographical
location. Collection of
blood, and urine sample
is required. High cost is
involved in blood and
urine tests.
Time consuming,
Accuracy is dependent
on sample collection,
storage, and equipment’s
available in laboratories.
Complete dataset
collected from only one
geographical location.
Prediction made only on
the basis of three
parameters. High cost
and time-consuming due
to laboratory tests.
The quality of ECG is
dependent on the
movement of the patient
during monitoring.
Tested on a small
dataset of only 10
patients.
-
Problem of class
imbalance in the dataset,
trained and tested on
very small dataset.

34 Disease Prediction using Machine Learning Rani et al.
https://t.me/med1917
(Table 1) co nt.....
Authors, Year,
and Citation
Isler et al.
Shahbazi et al.
Candelieri et al.
Zolfaghar et al.
Taslimitehrani
et al.
Leeza et al.
Technique Applied
Multi-layer
perceptron classifier
for predicting heart
fail severity
Generalized
Discriminant
Analysis (GDA) for
feature extraction
and KNN for Heart
fail severity
prediction
SVM hyper-solution
framework
Random Forest
algorithm.
CPXR (log)
classification
algorithm
Dictionary based
automated system
for severity
detection of Diabetic
Retinopathy (DR)
using Bag of Feature
(BoF) technique.
Dataset used
Dataset comprising
Heart Rate variability
Heart fail severity
Heart fail severity
Congestive heart
failure
Mayo clinic with
dataset comprising
5044 patients.
The dataset
containing 35126
images released by
Kaggle.
Performance
Accuracy of
96.43%
Reported
accuracy of 100%
Reported an
accuracy of
87.35%.
They reported an
accuracy of
87.12%.
93.7%
SVM achieves
accuracy of
98.3%,
Challenges
Small dataset, dependent
on results from
laboratory test reports
Small dataset, dependent
on results from
laboratory test reports.
Implemented on dataset
collected from a single
source. So, robustness of
model cannot be
evaluated.
Scope of improvement
in accuracy of the
model.
Implemented on dataset
collected from one
geographical location.
So, cannot claim for
robustness.
Implemented on dataset
collected from one
geographical location.
So, cannot claim for
robustness.
Similarly, DL techniques have been employed mainly for medical imaging data
such as Magnetic Resonance Imaging (MRI), Computer Tomography (CT) scans,
ultrasound images, and radiographs [11 - 28]. The capturing of medical images is
convenient for patients. The CT scan and MRI scans are expensive and
unaffordable to economically weaker sections of society. Also, these techniques
expose the patient to carcinogenic radiations. Thus, patients become prone to
cancer. Moreover, the availability of machinery used for CT scans and MRI is low
in remote and underdeveloped regions. Radiographs are low-cost, convenient, and
portable solutions for medical imaging. Also, X-ray machines are easily available
than CT scan and MRI machines.

Detection of Disease Severity Disease Prediction using Machine Learning 35
https://t.me/med1917
It has been observed from the review of literature that a number of DL techniques
have been employed for the automatic scanning of medical images for disease
diagnosis and severity detection. Although these techniques provide higher
accuracy, they have some limitations. Also, the efficacy of each technique is
highly dependent on the quality, type, and size of datasets used for training the
DL-based model. The comparative analysis of these techniques is shown in the
subsequent section in Table 2.
Table 2. Comparative analysis of deep learning techniques employed for disease severity detection.
Authors,
Year, and
Citation
Bodapati et al.
2021 [14]
Grover et al.
2018 [15],
Kermany et al.
2018 [16],
Chithra et al.
2020 [17]
Technique Applied Dataset Used Performance Challenges
Deep Neural Network
architecture with gated-
attention mechanism to
automatically diagnose
the Diabetic
Retinopathy.
DNN and adaptive
neuro-fuzzy inference
system with Support
vector regression
Deep learning based
diagnostic tool the AI
system.
Self-Adaptive Fractional
Crow Search-based
Deep Convolutional
Neural Network (AFC-
SVNN) model for the
severity detection and
infection level
identification of
Tuberculosis (TB)
disease.
Asia Pacific TeleOphthamology Society
(APTOS) dataset from
Kaggle.
5875 biomedical voice
measurements of 42
patients.
Screening of treatable
blinding retinal diseases
using 108,312 OCT
images for training and
1000 images for testing
Dataset comprising 75
images including 50
images of bacilli and 25
images of non-bacilli
classes.
Highest accuracy
of 97.32%.
Severity
prediction of
Parkinson’s
Disease with
accuracy of
81.6% using
DNN
Maximum
accuracy,
sensitivity and
specificity of
96.6%, 97.8%,
and 97.4%
respectively.
Best accuracy of
96.39% by using
KNN.
Faces the
problem of over
fitting and
multiple
channels.
Low accuracy,
conclusions based
on patients from
one geographical
location.
Requires dataset
of the same organ
for pre-training of
the system.
Implemented on a
small dataset,
thus less reliable
for practical use.

36 Disease Prediction using Machine Learning Rani et al.
https://t.me/med1917
(Table 2) co nt.....
Authors,
Year, and
Citation
Emrah Irmak
2021 [18]
Zhu et al. 2020
[19],
Rahimzadeh,
M. and A.
Attar [20],
2020
Li et al. [21],
2020
Chang et al.
[22], 2020
Li et al. [23],
2020
Technique Applied Dataset Used Performance Challenges
A fully automatic CNN
model for COVID-19
disease severity
assessment using chest
X-ray into four classes
as mild, moderate,
severe and critical.
A deep-learning CNN to
detect lung disease
severity of COVID-19
infection on portable
chest X-ray (CXR).
Xception and
ResNet50V2 based
neural network for
multiclass classification
for chest X-ray images
for three categories;
normal, pneumonia, and
COVID-19.
A deep convolutional
Siamese neural network
of disease severity in
Retinopathy of
Prematurity (ROP) in
retinal photographs and
Osteoarthritis.
Convolutional Siamese
network to identify knee
pain using MRI scans.
Siamese DNN to
estimate Psoriasis
Severity and locate the
skin lesion regions
based on skin lesion
images.
PhysioNet dataset
Publicly available dataset
comprising 131 CXRs
from 84 COVID-19
positive patients.
Dataset comprising 633
chest X-ray images.
Datasets containing 4861
images from 870 patients
for evaluating ROP and
10,012 images from 3021
patients for detecting
Osteoarthritis.
Dataset comprising MRI
slices from the two knees.
Dataset comprising 5,205
images from 1,787
Psoriasis patients.
The model
achieved average
accuracy of
95.52% .
Accuracy of
96.43%
Reported
average accuracy
of 91.4%.
Reported
receiving
operator
characteristic
area under the
curves of up to
0.90.
Achieved an
AUC of 0.808
Reported an
accuracy of
77.87%.
Focuses mainly
on COVID-19
and lacks assess
to other
pneumonias.
Small dataset,
and the study
does not identify
any correlation of
radiographic
severity with
clinical disease
severity.
Problem of data
imbalance is
observed.
Problem of data
imbalance is
observed.
Lack of
validation on
different imaging
datasets for full
spectrum of OA.
Lack of
visualization of
features.

Detection of Disease Severity Disease Prediction using Machine Learning 37
https://t.me/med1917
(Table 2) co nt.....
Authors,
Year, and
Citation
Zhang et al.
[24], 2020
Wang et al.
[26], 2021
Li et al. [28],
2020
Technique Applied Dataset Used Performance Challenges
Siamese structure based
deep learning method to
automatically detect
short-term lesion
changes in melanoma
screening.
Siamese network based
model with gated feature
fusion to simultaneously
assess the severity of
knee Osteoarthritis
(OA).
Implemented a
Convolutional Siamese
Neural Network that
calculates the
continuous radiography
pulmonary disease
severity score in
COVID-19.
Dataset collected from
Sydney Melanoma
Diagnostic Centre
(SMDC), Royal Prince
Alfred Hospital with
100,000 dermoscopy skin
lesion images.
OAI Dataset comprising
24319 and MOST dataset
comprising18634 single-
knee images.
Used dataset containing
161,590 anterior-posterior
images for training and
also applied transfer
learning on 314 frontal
Chest X-Rays (CXRs)
from COVID-19 patients.
The test dataset containing
154 and 113 CXRs from
different hospitals.
Achieved an
accuracy of
74.1%.
Not reported
Not reported
Scope of
improvement in
accuracy.
Lack of
validation of the
approach
employed.
Lack of
validation of the
approach
employed.
It is evident from the above discussion that there is a strong need to automate
disease severity detection for providing a quick, convenient, round-the-clock
available, and low-cost solution to health experts and the populace. This
motivated us to review the techniques proposed in the literature and present a
comparative analysis. In this chapter, we study the research works proposed from
2018 to 2022 to show the role of ML and DL techniques in severity detection. The
review encompasses the comparative analysis of ML and DL techniques
employed. It enlists the datasets used for severity detection. The paper also
highlights the drawbacks and limitations of the proposed works.
LITERATURE REVIEW
Severity Detection using Machine Learning
Yao et al. employed various machine learning algorithms such as predictive
Logistic Regression (PLR), Support Vector Machine (SVM), Random Forest
(RF), K-Nearest Neighbour (KNN), and AdaBoost to determine COVID-19
severity detection [9]. This study is performed on 75 severely infected COVID-19

38 Disease Prediction using Machine Learning Rani et al.
https://t.me/med1917
patients and 62 patients with mild symptoms. The data of patients were collected
from the Tongji hospital affiliated to Huazhong University of Science and
Technology. Each sample was analyzed to obtain 100 features comprising 8
clinical features such as sex, age, heart rate, respiratory rate, body temperature,
blood pressure etc., 76 parameters reported by blood tests, and 16 parameters
reported by urine tests. The missing entries of the dataset were filled either with
zero or the median of the normal range. Now, the dataset is divided into training
and testing sets in the ratio of 80% and 20% respectively. In the next step, the
authors perform T-test and identified 32 features that are statistically significant
for the disease severity. Further, the model applies Conservative Recursive
Feature Elimination (CRFE) strategy and eliminates 4 more redundant features.
These four features have strong inter-feature correlations while maintaining the
model performance. The authors reported the best prediction accuracy of 99.17%
by applying the Support Vector Machine (SVM) technique on the training dataset.
It achieved an accuracy of 81.48% when applied to the testing dataset. Further,
the authors claimed that the severely infected patients had a higher serum level of
neutrophil percentage and lower serum level of monocyte percentage and calcium
as compared to the patients with mild infection of COVID-19. Furthermore, the
parameters recorded by blood tests were found to be more significant in
identifying the severity than parameters recorded by urine tests. The authors also
claimed that males are at higher risk of severity if get infected by COVID-19.
However, the conclusions made by this study are important to predict the severity
of COVID-19, but there is a need to improve the reliability and robustness of the
model by using larger datasets from different regions. The study based on the
dataset of merely 75 patients from one geographical region of China cannot be
considered for making the generalized predictions.
In another research, Li Yan et al. proposed a three indices-based prognostic
prediction model that helps in differentiating critical cases from severe cases of
COVID-19. They also predicted the mortality risk of patients [10]. The authors
performed the experiments on a dataset comprising 375 patients from Tongji
Hospital in Wuhan, China. In the first step, they applied pre-processing
techniques and labelled the dataset with two classes’ viz. survival and death.
While pre-processing, the authors used padding of ‘-1’ to fill the missing values in
the dataset where clinical measures were incomplete. Now, the authors divided
the dataset into training and testing datasets in the ratio of 70% and 30%
respectively. They employed Multi-tree XGBoost model on the dataset. While
training the pre-set, the parameters were set as max depth value being 4, learning
rate as 0.2, the value of tress number of estimators as 150, and the regularization
parameter as 1. They pre-set the value of both ‘subsample’ and
‘colsample_bytree’ as 0.9 to avoid the problem of over-fitting. In the process of
feature selection, the algorithm selected three key features, i.e. lactic

Detection of Disease Severity Disease Prediction using Machine Learning 39
https://t.me/med1917
dehydrogenase (LDH), lymphocyte, and high-sensitivity C-reaction protein (hsCRP) from the pool of more than 300 features. This study precisely predicted
survival with more than 90% accuracy. The work proposed in this research
quantified the risk of death and may prove useful for patients. It is also a helpful
tool for doctors and predicted the severity and fatal development trend of the
disease. But, there is also a need to improve its reliability by training and testing it
with a vast dataset.
Tadesse et al. proposed the SVM-based ML model to detect Autonomic Nervous
System Dysfunction (ANSD) level that occurs due to Hand Foot and Mouth
Disease (HFMD) and tetanus [11]. The ANSD is the main cause of death for
HFMD and tetanus patients. The system was applied to physiological data of
patients collected using low-cost wearable sensors. It uses high pass filters
followed by a Gaussian filter to remove noises and movement artefacts from the
input data such as electrocardiogram waveform. Now, it computes gradients by
applying first-order derivatives to extract relevant features. These features can
capture more dynamic information in the time domain. The authors tested the
model on 60,373 samples that are extracted from the ECG signal of 74 HFMD
patients and ECG, PPG, and IP waveforms of 10 patients suffering from tetanus.
The dataset was collected from a hospital in Vietnam. The system employs the
SVM with the one-versus-all strategy to classify multiple classes from the HFMD
dataset. The obtained results are better for tetanus patients because mostly patients
of HFMD were children who are more likely to move. The movements of patients
while capturing the dataset lead to degradation in the data quality. This decreases
the classification performance of the system. However, the system reports high
accuracy but it requires manual encoding of features and needs automation to
improve the quality of the system.
Tripoliti et al. presented a review of ML-based models applied to predict the
presence of Heart Failure (HF), estimating its subtypes, determining the severity
of the disease, and predicting the presence of adverse results such as
destabilizations, re-hospitalizations, and mortality [12]. Based on the review of
literature, the authors concluded that K-NN and SVM techniques report high
accuracy for HF detection. Isler et al. proposed K-NN based technique and tested
it on the Physiobank database comprising ECG recordings of 54 normal and 29
heart patients [13]. They reported an accuracy of 96.39% with the value of K as 5
and 7 in the K-NN algorithm. Similarly, Liu et al. presented the SVM-based
model and tested it on PhysioNet dataset collected from 30 normal and 17 heart
patients. They used the short-(Heart Rate variability) HRV measures as predictor
feature, and reported an accuracy of 100%.

40 Disease Prediction using Machine Learning Rani et al.
https://t.me/med1917
Another team of researchers, Isler et al. proposed a multi-layer perceptron
classifier that utilizes short-term HRV measures as input and classifies HF into
systolic and diastolic CHF with an accuracy of 96.43%. Similarly, Shahbazi et al.
performed feature extraction with Generalized Discriminant Analysis (GDA) on
long term HRV measures feature to determine HF severity estimation using K-NN
with an accuracy of 100%. But, they evaluated the performance of their technique
on a very small and unbalanced dataset. Therefore, generalization of the results is
not possible. Candelieri et al. implemented the SVM hyper-solution framework
that employs genetic algorithm based meta-heuristic search for the most reliable
hyper-classifier for destabilization and gave an accuracy of 87.35%. Similarly,
Zolfaghar et al. predicted the risk of readmission for Chronic Heart Failure (CHF)
within a period of 30-days using the Random Forest algorithm. They reported an
accuracy of 87.12% on the dataset comprising features such as sociodemographic, lab tests, discharge disposition, and medical comorbidity.
Next, Taslimitehrani et al. employed the CPXR (log) classification algorithm that
uses a pattern as logical characterization of a subgroup of data and a regression
model to characterize the relationship between predictor and response for data of
that group. They performed experiments on electronic health records of the Mayo
clinic with the dataset comprising 5044 patients. The technique proposed in this
research outperformed the decision trees, Random Forest, SVM, AdaBoost, and
logistic regression with an accuracy of 93.7% for prediction of mortality rate in
the age of 1 year, 83% for the age of 2 years, and 78.6% for the age of 5 years.
Leeza et al. proposed a dictionary based automated system for severity detection
of Diabetic Retinopathy (DR) using a Bag of Feature (BoF) technique [14]. This
approach did not use any pre-processing and post-processing steps. BoF allows
image collection and also helps in identifying visual patterns of the whole image
collection. The authors applied speed up robust features algorithm and histogram
of oriented gradients to compute the descriptive features of retinal images. The
algorithm uses SURF descriptor for feature extraction. The SURF descriptor is
invariant to image rotation, illumination changes and scale. BoF uses coding and
pooling to preserve valuable information that can be lost during quantisation of
visual words and loss of spatial information. The system classified the severity
level of DR into five classes using radial basis kernel SVM and Artificial Neural
Networks (ANN) to classify input images. The four-layered ANN with scaled
conjugate back propagation technique employs gradient descent to reduce the
mean squared error. SVM achieves 98.3% accuracy, 95.92% sensitivity and
98.9% specificity and as 95.74% PPV whereas ANN achieves an accuracy of
92.92%, specificity of 95.97%, sensitivity of 83.83% and PPV of 83.82% on the
test dataset. The dataset containing 35126 images is released by Kaggle. It is
apparent from the results that the proposed system achieved better results using

Detection of Disease Severity Disease Prediction using Machine Learning 41
https://t.me/med1917
visual features with SVM. This proposed system with SVM algorithm improved
the accuracy by 1% over the previous DNN based approach.
This is apparent from the above discussion that ML techniques such as KNN, RF,
SVM, AdaBoost, logistic regression, and ANN have been employed on the
dataset collected from hospitals and publicly available datasets. These techniques
reported the accuracy in the range of 78% to 100% on various datasets for the
detection of severity of diseases such as heart failure, tetnus, hands and mouth
disease, COVID-19, and diabetes. The comparative analysis of these techniques
has been presented in Table 1. It is evident from the analysis presented in Table 1
that mostly techniques have been applied on a small dataset collected from a
specific geographical location. Moreover, blood tests and urine tests are initial
requirements for employing these techniques for disease severity detection. Thus,
these techniques are not the alternative for laboratory tests, but used as a tool for
reporting and data analysis for predicting disease severity. The comparative
analysis of the existing techniques is discussed in Table 1.
Inference deduced: From the analysis of machine learning techniques employed
so far, we conclude that there is a huge scope to implement these techniques on
dataset collected from various geographical locations, and at various stages of a
disease for improving the reliability of robustness of these techniques. Also, the
severity detection is applied for a very less number of diseases. This raises the
demand for extending the applications of ML techniques for severity prediction of
more diseases. Further, we conclude that there is a scope to integrate the ML
techniques with IoT for data collection, and communication. Also, the integrated
model of IoT, and ML can be available at a cloud storage for hassle free data
sharing, processing, round the clock availability, and location independent
accessing of prediction results. The sensor based data collection may reduce the
number of laboratory tests required for prediction of severity of a disease.
Severity Detection using Deep Learning
Bodapati et al. proposed a composite Deep Neural Network (DNN) architecture
with gated-attention mechanism to automatically diagnose the Diabetic
Retinopathy (DR) disease [14]. The gated-attention mechanism allowed the
system to emphasize on the input descriptors that represent images of retinal
lesion portions. The system used two pre-trained models such as VGG16 and
Xception to extract important features of color fundus images. The authors used
the Asia Pacific Tele-Ophthamology Society (APTOS) dataset from Kaggle. The
high dimensional features obtained from the above models may cause overfitting
when used as input. Therefore, these features are reduced using the proposed
spatial pooling method. The authors worked upon the idea that the use of multiple

42 Disease Prediction using Machine Learning Rani et al.
https://t.me/med1917
image representations as input will complement each other and improve the
system’s performance. They implemented the composite DNN model with two
channels that accept two different representations of the retinal representations,
process them and learn the information about lesion present in those images. The
authors compared the proposed composite DNN with other ML methods such as
linear regression, KNN, Decision Tree (DT), SVM and Multi-Layer Perceptron
(MLP) and showed that DNN outperforms all these methods with an accuracy of
97.32%. The major limitation of this work is that it has poor performance in
higher severity level prediction due to inadequate samples.
Grover et al. proposed a deep learning-based methodology to determine disease
severity [15]. They worked on UCI’s Parkinson’s Tele-monitoring Voice Data set
of patients. The dataset contains 5875 biomedical voice measurements of 42
patients. The authors employed the classifier containing 16 units in the input
layer, 10, 20, and 10 neurons in each of the hidden layers respectively. The
technique utilized the Total Unique Parkinson’s Disease Rating Scale (UPDRS),
Motor UPDRS, and 16 biomedical voice measures the features for prediction.
Based on the values of UPDRS and motor UPDRS, the system classifies the
dataset into severe and non-severe classes. The system attained an accuracy of
62.733% with a total UPDRS score and an accuracy of 81.667% with the motor
UPDRS score. The authors concluded that motor UPDRS is a better feature for
classifying Parkinson’s disease as compared to total UPDRS. Based on the
comparative analysis of Deep Neural Network (DNN) and adaptive neuro-fuzzy
inference system with Support vector regression, the authors claimed the
supremacy of the DNN approach.
Next, Kermany et al. developed a deep learning-based diagnostic tool for
screening of patients with common treatable blinding retinal diseases [16]. The
authors proposed a transfer learning-based algorithm to meet the challenge of a
small dataset. They tested their algorithm for Optical Coherence Tomography
(OCT) images of the retina and the pediatric chest radiograph to verify its
generalizability. They used the dataset comprising 108,312 OCT images for
training and 1000 images for testing the AI system. The system attained its
maximum accuracy, sensitivity, and specificity of 96.6%, 97.8%, and 97.4%
respectively after 100 epochs. Further, training the pre-trained system using the
dataset comprising 1000 images reported an accuracy of 93.4%. This shows the
importance of utilizing transfer learning.
Chithra et al. proposed a self-Adaptive Fractional Crow Search-based Deep
Convolutional Neural Network (AFC-SVNN) model for the severity detection and
identification of infection level of Tuberculosis (TB) disease [17]. They followed
a five-step process that includes preprocessing, segmentation, feature extraction,
Соседние файлы в папке Библиотека им академика М.И. Перельмана
