APPLYING MACHINE LEARNING TECHNIQUES TO FIND IMPORTANT ATTRIBUTES FOR HEART FAILURE SEVERITY ASSESSMENT

International Journal of Computer Science, Engineering and Applications (IJCSEA) Vol. 7, No. 5, October 2017
DOI: 10.5121/ijcsea.2017.7501 1
APPLYING MACHINE LEARNING TECHNIQUES TO FIND
IMPORTANT ATTRIBUTES FOR HEART FAILURE
SEVERITY ASSESSMENT
Puram Surya Prudvi 1
and Ershad Sharifahmadian 2
1
Department of Computer Engineering, Masters in Computer Engineering, University of
Houston Clear Lake, Houston, Texas, USA
2
Department of Computer Engineering, Visiting Asst Professor, University of Houston
Clear Lake, Houston, Texas, USA
ABSTRACT
The diagnosis of heart disease depends mostly on the combination of clinical and pathological data. It
leads to the quality of medical care provided for the patient. In this paper, three machine learning (ML)
techniques −Classification and Regression tree (CART), Neural Networks (NN), and Support vector
machine (SVM)− are utilized to find the best attributes for estimating the severity of heart failure. The data
is collected from three different resources, then each input attribute used for assessing the severity of heart
failure is analyzed individually after implementing the machine learning techniques. Finally, the most
important supportive attributes are presented in this paper by which medical staffs can identify heart
failure severity fast and more accurately. In fact, by screening important attributes, clinicians can make
better decision about right treatment procedures or preventive actions that reduce risk of heart attacks.
KEYWORDS
Congestive Heart Failure (CHF), Decision support system, Heart failure severity, Machine learning, Risk
classification
1. INTRODUCTION
Heart failure (HF) occurs due to an insufficient supply of blood from the heart. To meet the
general body needs a certain amount of blood is necessary. Breathlessness, insufficient sleep,
excessive tiredness, swelling of legs are some of the symptoms of the heart failure. Heart failure
is not same as a heart attack which is caused due to damage of the heart muscle. Some of the
common causes of heart failure are heart attack, high blood pressure, excess alcohol consumption,
drugs consumption, cigarette smoking, atrial fibrillation etc. High output of blood also causes
heart failure. When the amount of blood pumped by the heart is greater than the typical amount of
blood and the heart is not able to keep up, and then high-output heart failure will occur, which can
be termed as Congestive heart failure (CHF). Person affected with CHF usually has substantial
symptoms such as shortness of breath and chest pain.
Heart failure management includes the perpetuation of life, reduction of symptoms and being
more activity. The maintenance of heart failure is very important for the person affected with HF.
The Heart failure severity assessment and type prediction are important when the patient
condition is analyzed [1]. Thus, for the severity assessment, several symptoms from patients
should be observed. The main goal of this study is to find important attributes by which the
severity assessment of heart failure is better identified. This goal can be achieved based on
classification of data using the machine learning techniques. The data with different attributes in
different experiments are given to ML techniques. Then, based on the results of classification (i.e.

2
output of ML techniques) important attributes are selected. The severity assessment is
exceptionally helpful, for instance, in remote healthcare monitoring. In this paper, the severity
assessment is done using three machine learning techniques including CART (classification and
regression tree), SVM (support vector machine), and NN (neural networks). Then, the
performance of ML techniques is evaluated for each attribute (i.e. symptom) separately. In this
study, the results are evaluated such that the attribute which helps to assess the HF severity more
accurately is considered as one of the important supportive attributes.
2. RELATED WORK
In the previous works several approaches for assessing severity of Heart Failure, other diseases
and various Machine learning approaches for classification are proposed.
Gabriele Guidi, Maria Chiara Pettenati, Paolo Melillo et.al. [1] proposed a clinical decision
support system for assessing the severity. In the support system, a management interface is built
for the heart failure type prediction and severity assessment. To implement the smart functions
machine learning techniques are implemented. P. Melillo, N. De Luca, M. Bracale, and L.
Pecchia et.al. [5] proposed Classification tree based risk assessment for separating higher risk
patients from lower risk patients using of long term heart rate variability measures. T.John Peter,
K. Somasundaram et.al. [9] used Naïve Bayes, K-Nearest Neighbour, Decision Tree for
prediction of risk in heart failure patients. Kavetha.BV, Venu Gopala Krishnan.J, et.al [10] used
CVPartition method for classifying, deciding and detecting Maligant and Benign in
mammorgams. Amiya Halder, Oyendrila Dobe et.al [17] explained about Fuzzy feature selection
and support vector machine for detecting Tumor in Brain MRI.
The present work describes about three machine learning approaches for assessing the severity of
heart failure and the important attributes for assessing severity are concluded by several insights
into the data.
3. METHOD
To analyze the importance of supportive attributes in estimating the severity of heart failure, the
outcome of ML techniques is considered. The ML techniques classify patients based on the
severity level of the heart failure as i) mild, ii) moderate and iii) severe. The general block
diagram of HF severity assessment is given in the figure 1.
Classification and Regression Trees is an order strategy which utilizes chronicled information to
build purported decision trees. Decision trees are then used to group additional information. To
utilize CART, we need to know number of classes from the earlier [2], [3].
we are going to discuss the three machine learning techniques and their role in classification:

3
a. Decision trees arrange occurrence by sorting them down the tree from the root to some
leaf node, which gives the arrangement of the occurrence. Every node in the tree
determines a trial of some trait of the occurrence and each branch downward from that
node relates to one of the conceivable qualities for the given attribute. An occurrence is
ordered by beginning at the root node of the tree, testing the property determined by this
node, and then moving down the tree limb comparing to the estimation of the
characteristic. This procedure is then continual for the sub- tree rooted at the new node
[4].
b. Support vector machine (SVM) forms a model that allocates new examples to
Classification or regression analysis for a given set of training data sets [5]. The SVM
model is an illustration of the training examples with space as a plane and the data as
points and are cleverly mapped, separated by a clear gap to project that they belong to
two distinct categories
c. Neural networks learning method is a computational approach based on a rough analogy
of artificial neural networks which are an enormous collection of neural units exhibiting
the exact same way the brain solves problems with the help of large clusters of biological
neurons [4]. The feed-forward neural network is also implemented. By varying the
hidden neurons from 2 to 8. The best configuration is 8 neurons for the severity
assessment.
4. RESULTS
In this paper, we used the heart disease data from three resources. The database I is an
anonymized database of HF patients, with varying severity degrees, all treated by the Cardiology
Department at the St. Maria Nuova Hospital, Florence, Italy, in the period 2001–2008 [1]. This
database consists of 136 records from 90 patients, including baseline and follow-up data (when
available). At the time of the data collection, the specialist physician provided the mentioned HF
severity assessment in the desired three levels: i) mild, ii) moderate, and iii) severe, which was
stored in the database. 12 variables (i.e. attributes) in this database that are used as input for the
machine learning techniques are the following:
1) Anamnestic data: age, gender, and New York Heart Association (NYHA) class.
2) Instrumental data: weight, systolic blood pressure, diastolic blood pressure, EF (Ejection
Fraction), BNP (Brain natriuretic peptide), heart rate, ECG parameters (atrial fibrillation
true/false, left bundle branch block true/false, and ventricular tachycardia true/false).
The database II is machine learning repository of UCI [6] which was collected from the
Cleveland Clinic Foundation.
We have 303 instances of which 164 instances belonged to the healthy cases and 139 instances
belonged to the heart disease. 14 clinical features have been recorded for each instance. The table
I shows the 14 clinical features and their description.

4
Table I- Clinical features and their description
The database III is anonymized data collected from 246 patients with 14 attributes such as age,
sex, Dyspnea, smoking, dust, Respiratory frequency, Inhale and exhale time, ECG ST
segmentation, Heart rate, peripheral capillary oxygen saturation (spo2), systolic and diastolic
blood pressures. The data is collected from Siddhartha government medical hospital, Vijayawada,
India. The data provides an overview of prevalence of Chronic Obstructive Pulmonary Disease
(COPD) in Vijayawada, India. It provides insights into the mortality, morbidity and etiological
determination of COPD and emphasis in understanding the multidimensional nature of problem.
All the data is normalized eliminating redundant data (for example, storing the same data in more
than one table) and ensuring data dependencies make sense.
Using the machine learning techniques, we evaluate the performance of each attribute in
assessing the severity of HF. First, input data is considered from three resources and then the
common attributes in the three datasets are considered. Each attribute is used for the ML
techniques along with the common attributes, and the performance in assessing the HF severity is
observed.
Three machine learning techniques are applied and the corresponding results are as presented:
We extract the common attributes from the three datasets. The common attributes are age, sex,
Heart rate, blood pressure. Then, we examine the performance of each method in assessing the
severity of heart failure based on each supportive attribute. First, Classification and regression
tree (CART) is examined with the common attributes. Then each supportive attribute from each

5
dataset is added to the common attributes and results are studied separately. Later the support
vector machine (SVM), and NN are inspected in the same way. The accuracy, precision, and
sensitivity are calculated for each method individually. Each test is done using MATLAB
R2016b.
The accuracy of the three ML techniques using the three datasets are shown in Figures 2, 3,4. For
database I, common attributes are considered and their role is evaluated in assessing the severity
of the heart failure. Then, supportive attributes such as BNP, ECG parameters, NYHA class, EF
rate, Weight are used with the common attributes. Then, the accuracy of each ML techniques is
calculated. Among all the supportive attributes from the database I, BNP, NYHA class, ECG
parameters provide higher accuracy comparing other input attributes as shown in the figure 2.
Fig 2: Graph showing the accuracy of the machine learning techniques for the database I
Fig 3: Graph showing the accuracy of the machine learning techniques for the database II.
Fig 4: Graph showing the accuracy of the machine learning techniques for the database III.

6
Next, the common attributes from the database II are considered, then the supportive attributes
such as Cholesterol, Chest pain, ECG parameters, Exercise induced angina, fasting blood sugar,
ST segment, Thalassemia are considered. Among all the supportive attributes, Cholesterol, Chest
pain, ECG parameters and ST segment provide higher accuracy as shown in the figure 3. Last,
ML techniques are applied on the database III collected from Siddhartha government medical
hospital, Vijayawada, India. After that, the supportive attributes are identified and the role of each
attribute in assessing the severity of the heart failure is observed. Of all the supportive attributes,
the Respiratory Frequency (RF), spo2, ECG parameters, and Dyspnea provide higher accuracy as
shown in the figure 4.
5. DISCUSSION
The selection of the important supportive attributes among all attributes in the heart failure
severity assessment is done not only based on the accuracy but also based on the precision and
sensitivity. The rate of precision is observed for each method and evaluated for each level of
severity (i.e. Mild, Moderate and Severe). The precision for each level is tested with different
supportive attributes along with the common attributes. The figure 5 demonstrates the variation in
precision for each class (i.e. Mild, Moderate and Severe) when the CART is implemented for
different supportive attributes.
Fig 5: Graph showing precision values of all classes with respective to CART technique using the different
supportive attributes along with common attributes.
The figure 6 explains the sensitivity of each class when the CART is implemented for different
supportive attributes. In medical diagnosis, test sensitivity is the ability of a test to correctly
identify those with the disease. In other words, sensitivity is the extent to which the true positives
are not overlooked. When we observe the accuracy, precision and sensitivity rates of different
supportive attributes while implementing CART we noticed that BNP, chest pain, smoking, ECG
parameters, and dyspnea exhibited the better results. Therefore, we can select those as important
attributes in severity assessment of heart failure during the CART implementation.
Fig 6: Graph showing sensitivity values of all classes with respective to CART technique using the
different supportive attributes along with common attributes.

7
In the same way, SVM and NN are used. The supportive attributes and the common attributes are
evaluated. Figures 7, and 8 show the precision, sensitivity for each level of severity (i.e. class)
when the SVM is implemented. After implementing the SVM, we find out that BNP, smoking,
ECG parameters, cholesterol, chest pain, dyspnea are important supportive attributes.
Fig 7: Graph showing precision values of all classes with respective to SVM technique using the different
supportive attributes along with common attributes
Fig 8: Graph showing sensitivity values of all classes with respective to SVM technique using the different
Figures 9 and 10 demonstrate the precision, and sensitivity for each class when the NN is
implemented. After NN implementation, we conclude that ECG parameters, cholesterol, Chest
pain, smoking, dyspnea are the important supportive attributes as those performed well
comparing to other attributes. CART produced satisfactory results in severity assessment if
compared with other studies that assess HF severity such as [1] that classify HF patients in three
groups mild, moderate, and severe. As shown in the graphs, the accuracy, precision, and
sensitivity are calculated for each supportive attribute individually.
Fig 9: Graph showing precision values of all classes with respective to NN technique using the different

8
Based on results, CART outperforms well with the accuracy of 84.4%. For different machine
learning techniques, different supportive attributes are considered. the SVM has the average
accuracy of 76%. The neural network has the average accuracy of 78%.
Fig 10: Graph showing sensitivity values of all classes with respective to NN technique using the different
5. CONCLUSION
T This work identifies main attributes for fast and more accurate assessment of the heart failure
severity in patients and the status of the disease. After evaluating selected clinical observations
from patients (i.e. Main attributes), physicians and other health professionals can better choose
right treatment procedures or preventive actions that reduce risk of heart attacks. First, we
selected the common clinical attributes such as age, sex, heart rate etc. Three datasets were used.
Those databases were selected to comprehensively evaluate broad ranges of clinical parameters
which influence the heart failure severity assessment. The three machine learning techniques were
implemented to identify the main supportive attributes. Different ml techniques were used to
show that identified main attributes are independent from ml techniques, in other words, the
changing classification method (i.e. ML technique) will not significantly affect the main
supportive attributes. Later we evaluated the performance of each technique by adding different
clinical attributes individually to the common attributes. After examining the performance of the
ml techniques, main clinical attributes were identified as important supportive attribute for each
technique. After cart implementation, we notice that BNP, chest pain, smoking, ECG parameters,
and dyspnea exhibited the better results. After implementing the SVM, we find out that BNP,
smoking, ECG parameters, cholesterol, chest pain, dyspnea are important supportive attributes.
After NN implementation, we conclude that ECG parameters, cholesterol, chest pain, smoking,
dyspnea are the important supportive attributes. Cart outperforms well with the accuracy of
84.4%. The SVM has the average accuracy of 76%. The neural network has the average accuracy
of 78%. Among all the methods, cart technique provided the better results in severity assessment
of heart failure with the accuracy of 84.4%.
ACKNOWLEDMENT
The authors would like to thank Dr. Gabriele Guidi, Robert Detrano, Dr. M.P.P. Bala
Narasimhulu for their support and providing database for this project.
REFERENCES
[1] Gabriele Guidi, Maria Chiara Pettenati, Paolo Melillo, “A Machine Learning System to Improve
Heart Failure Patient Assistance,” IEEE J. biomedical and health informatics, vol. 18, no. 6, pp. 1750-
1756, Nov 2014

9
[2] Priyanka Pandey,Amita Jain, “A Comparative study of Classification techniques: Support vector
Machine & Decision Tress”, . International Conference on Computing for Sustainable Global
Development, pp.3620-3624, 2016.
[3] Roman Timofeev, “Classification and Regression Tress. Theory and Applications”, Cent. Of Appl.
Statis. And Econ., Humboldt Univ., Berlin, Dec.20,2004.
[4] s [Online].
[5] P. Melillo, N. De Luca, M. Bracale, and L. Pecchia, “Classification tree for risk assessment in patients
suffering from congestive heart failure via long-term heart rate variability,” IEEE J. Biomed. Health
Inform., vol. 17, pp. 727–733, May 2013.
[6] UCI Machine Learning Repository [homepage on the Internet]. Arlington: The Association; 2006
[updated 1996 Dec 3;cited 2011 Feb 2].Available from:
http://archive.ics.uci.edu/ml/datasets/Heart+Disease.
[7] M. Sokolova and G. Lapalme, “A systematic analysis of performance measures for classification
tasks,” Inf. Process. Manage., vol. 45, no. 4, pp. 427–437, Jul. 2009.
[8] Chun-Fu Lin and Sheng-De Wang, “Fuzzy Support Vector Machines”, IEEE Transactions on Neural
Networks, Vol.2, No.2, March 2002.
[9] T.John Peter, K. Somasundaram, “An empirical study on prediction of heart disease using
classification data mining techniques”, IEEE-International Conference On Advances In Engineering,
Science And Management (ICAESM -2012) March 30, pp.514-518, 2012.
[10] Kavetha.BV, Venu Gopala Krishnan.J, “Decide,Detect And Classify Benign And Maligant in
Mammorgams using CV-Partitioning Method”, ARPN Journal of Engineering and Applied
Sciences.Vol.10,No.10, June 2015.
[11] AH Chen, SY Huang, PS Cheng, EJ Lin, “HDPS: Heart Disease Prediction System. Computing in
Cardiology”, Vol.38, pp. 557-560, 2011.
[12] Ryusuke Hata, Kazuyuki Muras, “Multi-valued Autoencoders for Multi-Valued Neural Networks”,
IEEE Trans. International Joint Conference on Neural Networks, pp.4412-4417, 2016.
[13] Atul Kumar Pandey, Prabhat Pandey, K.L. Jaiswal, Ashish Kumar Sen, “A Heart Disease Prediction
Model using Decision Tree. IOSR Journal of Computer Engineering”, Vol.12, Issue 6, pp. 83-86, Aug
2013.
[14] Seyed Saleh Mohaseni Matiar Mohamadyrai, “Heart Arrhythmias Classification via a Sequential
Classifier Using Neural Network, Principal Component Analysis and Heart Rate Variation”, IEEE
Trans.8th International Conference on Intelligent Systems, pp.715-722, 2016.
[15] Udaya Sampath K. Perara Miriya Thanthrige, Jagath Samarabandu Xianbin Wang, “Machine
Learning Techniques for Intrusion Detection on Public Dataset”, IEEE Trans. Canadian Conference
on Electrical and Computer Engineering, 2016.
[16] Fabricio Aparecido Breve, Daniel Carlos Guimaraes Pedronette, “Combined Unsupervised and Semi-
Supervised Learning for Data Classification”, IEEE Trans. International WorkShop on Machine
Learning for Signal Procession, sep.2016.
[17] Amiya Halder, Oyendrila Dobe, “Detection of Tumor in Brain MRI Using Fuzzy Feature Selection
and Support Vector Machine”, International Conference on Advances in Computing,
Communications and Informatics, pp.21-24, sep.2016.
[18] B.Venkatalakshmi, M.V Shivsankar, “Heart Disease Diagnosis Using Predictive Data mining”,
ICIET’14, Volume 3, Issue 3, March 2014.

APPLYING MACHINE LEARNING TECHNIQUES TO FIND IMPORTANT ATTRIBUTES FOR HEART FAILURE SEVERITY ASSESSMENT

More Related Content

What's hot (20)

Similar to APPLYING MACHINE LEARNING TECHNIQUES TO FIND IMPORTANT ATTRIBUTES FOR HEART FAILURE SEVERITY ASSESSMENT (20)

Recently uploaded (20)

APPLYING MACHINE LEARNING TECHNIQUES TO FIND IMPORTANT ATTRIBUTES FOR HEART FAILURE SEVERITY ASSESSMENT