Deep Neural Network based Heart Disease Prediction

Bixi Zhang

School of Physics, Nanjing University, Jiangsu Nanjing, 210093, China

[email protected]

Abstract. Since the 21st century, cardiovascular disease (CVD), which has a
high rate of morbidity and mortality, has become one of the most prevalent and
fatal diseases. Therefore, machine-learning and deep-learning-based heart
disease prediction models hold substantial research value in the medical field
with extensive potential applications and great significance. This article aims to
establish a prediction model, utilizing the Deep Neural Network (DNN)
algorithm, to cope with latent hazards associated with heart disease. A dataset
from the Kaggle database was used for this approach. To assess the performance
of the prediction model, evaluation indices including accuracy, recall_0, recall_1,
and AUC, etc. were calculated for DNN and five comparative machine-learning
algorithms. DNN eventually achieved outstanding performance with an accuracy
of 0.76, recall_1 rate of 0.77, and AUC value of 0.84, respectively. The study
concluded that DNN showed better average performance compared to other
machine learning algorithms. The result could serve as an auxiliary strategy for
heart disease diagnoses.

Keywords: Heart Disease, Deep Neural Network, Machine learning, Feature

selection, Imbalance learning.

1 Introduction

Since the 21st century, cardiovascular diseases (CVDs) have consistently been one of
the deadliest illnesses, leading to massive fatalities worldwide. Heart disease is the
critical branch of cardiovascular disease. According to Megan Lindstrom et al., the
global all-age mortality rate from heart disease in 2021 is 258.8 per 100,000 people,
and there is a prevalence of up to 7852.0 per 100,000 people [1]. With the increase in
the modern standard of living, there is a growing emphasis on improving quality of life
due to socio-economic and cultural prosperity. Heart disease could pose a serious
potential threat to health, resulting in a dramatic decline in quality of life when it occurs
to individuals. Therefore, establishing a systematic strategy to predict heart disease
would allow earlier treatments for potential patients with medication or surgery, thus
reducing its serious consequences.
In recent decades, owing to advancements in computer processing power, AI has
made significant progress in disease prediction, particularly in heart disease, skin
disease, cancer, etc. [2]. Meanwhile, the open-source databases provided by the
Internet facilitate the training of artificial intelligence models. Numerous machine

© The Author(s) 2024

Y. Wang (ed.), Proceedings of the 2024 International Conference on Artificial Intelligence and Communication
(ICAIC 2024), Advances in Intelligent Systems Research 185,
https://doi.org/10.2991/978-94-6463-512-6_70
668 B. Zhang

learning and deep learning models for heart disease prediction based on cardiology
datasets such as Cleveland and IEEE have already been established and exhibited [3].
These models are relatively well-performing.
Electrocardiograms (ECGs) can provide valuable information when using artificial
intelligence to predict heart disease [4]. Y. M. Ayano et al. provided a targeted review
of ECG-based heart disease prediction focusing on Interpretable Machine Learning
(IML) [5]. E. H. A.-E. Rabie Ahmed et al. overviewed the diagnosis and classification
of heart failure. Machine learning strategies comprising supervised, unsupervised, and
reinforcement learning were exploited [6]. J. Botros et al. realized a CNN-based
classification method for ECGs and applied the model to heart failure prediction,
achieving 99.31% accuracy, 99.50% sensitivity, and 99.11% specificity [7]. I. M. El-
Hasnony et al. utilized the Cleveland database and implemented five active learning
strategies to train and predict heart disease in a manner different from traditional
supervised learning. However, the final performance was unsatisfactory, with an
accuracy of 57.4% ± 4% and an F-score of 62.2% ± 3.6% [8]. C.-Y. Guo et al.
developed a prediction model for heart failure using four machine-learning models.
They also applied, for instance, the KNN clustering algorithm to classify the samples
and enhance the performance of their models. Despite its high level of
professionalism, the dataset from the JHS database, which was applied for machine
learning contained an excessive number of variables (over 100), which hindered the
practical application of their models [9].
The majority of previous research centered their view on specialized medical data,
such as ECG data, resting blood pressure (restbps), plasma cholesterol level (chol),
fasting blood sugar (fbs), etc., which can be difficult to obtain, constituting an
obstacle to the practical application. Moreover, variables incorporated in the datasets
applied by these studies are generally involved, making the algorithm intricate and
slow. Overall, it’s crucial to develop a prediction model for heart disease that is
accessible for daily use.
This paper intends to develop a prediction model for heart disease with accessible
variables, to streamline heart disease prediction in daily life. A novel dataset (2022)
from Kaggle with 17 variables was employed for this approach. The model was based
on Deep Neural Network (DNN) algorithm, and five machine learning algorithms
were analyzed for comparison. Feature screening and unbalanced strategies were
adopted for performance optimization. The evaluation results were presented
separately in section 3.

2 Dataset And Methods

2.1 Dataset
The Dataset in this research was chosen from the Kaggle database. 319795 instances
were included in the total. A yes/no approach was used to record whether or not there
had been a cardiac heart disease (CHD or MI). Seventeen relevant variables, whose
specific meanings and values are shown in Table 1, were concerned. It should be
noted that the number of instances counted as “yes” (Potential patient for heart
 Deep Neural Network based Heart Disease Prediction 669

disease) in the dataset was 27.4k, merely 9% of the total samples. While those
recorded as “no” (Not a potential patient for heart disease) were 292k, about 91% of
the total samples. The dataset consequently presented a tremendous imbalance
essence.
Before training, this paper used the Label Encoder tactic to convert binary and
discrete “object” values into “numeric” values. Subsequently, the dataset was split
into two segments, with a ratio of 7:3 for training and testing, respectively.

Table 1. Introduction of Variables

Variables Contents Type Values

BMI Body Mass Index discrete 23, 24, 25
Have you ever smoked at least 100 cigarettes in your entire
Smoking binary 0, 1
life
AlcoholDrinking Are you a heavy drinker binary 0, 1
Stroke Have you had a stroke or not binary 0, 1
For how many days during the past 30 do you have physical
PhysicalHealth discrete 1--30
health problems
For how many days during the past 30 do you have mental
MentalHealth discrete 1--30
health problems
DiffWalking Do you have serious difficulty walking binary 0, 1
Sex Your gender binary 0, 1
AgeCategory Your age discrete 1, 2, 3
Race Your race discrete 1, 2, 3, 4
Diabetic You had diabetes or not discrete 1, 2, 3, 4
PhysicalActivity Were you doing regular exercise during the past 30 days binary 0, 1
GenHealth Your general health condition discrete 1, 2, 3, 4
SleepTime Your average sleep hours discrete 1, 2, 3
Asthma You had asthma or not binary 0, 1
KidneyDisease You had kidney disease or not binary 0, 1
SkinCancer You had kidney disease or not binary 0, 1

2.2 Methods

Deep Neural Network. Fully connected deep neural networks (DNNs) are neural
network algorithms based on multi-layer perceptrons (MLPs). They offer higher
accuracy and more powerful nonlinear fitting capabilities than traditional machine
learning algorithms. DNNs are capable of creating extremely accurate decision
boundaries in binary classification problems. While fully connected deep neural
networks are prone to overfitting and require a long training period, they could still be
a good choice for the prediction algorithm in this study due to simple variables and
the small data volume of the selected dataset.
670 B. Zhang

The structure of the neural network is depicted in Fig. 1. This article designed a
network with 2 hidden layers containing 64 and 32 neurons, respectively. Unlike
traditional DNNs, the neural network in this paper was deliberately designed as "wide
and shallow". The structure aimed at striking a balance between comprehensive
extraction of features and reducing overfitting. To further reduce overfitting, a ReLU
function (refer to Fig. 2) with sparse activation property were chosen as the hidden
layer activation function. The Sigmoid function S (x) was used as the output layer
function:

The loss function chosen for the neural network was the cross-entropy
function J (p, y):

Whereas y represents the real value (in this study, only 1 or 0 was taken,
representing diseased and not diseased, respectively), p represents the prediction
probability generated from the neural network (a value between 0 and 1). The Adam
optimizer was used as a gradient descent rule [10].

Fig. 1. Deep Neural Network(Photo/Picture credit : Original )

 Deep Neural Network based Heart Disease Prediction 671

Fig. 2. ReLU Function(Photo/Picture credit : Original )

Comparative Machine Learning Methods. To showcase the effectiveness of fully

connected deep neural networks, the study selected five machine learning models to
compare their performance with the former algorithm. For those five machine
learning models applied, their features and advantages are listed below:
Logistic Regression is fast classification with a small computation volume for
quick results. Excellent fit for linear binary classification problems.
Decision Tree is white-box and highly interpretable. Skilled at handling non-linear
problems, but susceptible to overfitting.
In contrast to the decision tree model, Random Forest is highly resistant to
interference and provides an effective method to mitigate errors when dealing with
unbalanced datasets. It significantly outperforms the decision tree in unbalanced
datasets.
XGB and GBDT are integrated learning methods with strong robustness and have
performed well in AI algorithm competitions in recent years. This study selected them
for comparison.

Evaluation Indices. To evaluate the prediction model's performance, mainly four

indices were applied, including accuracy, recall_0, recall_1, and AUC. The definitions
are given below, respectively:
Accuracy: The percentage of samples that were predicted correctly out of the total
number of predictions.
Recall_0: Accuracy among negative samples.
Recall_1: Accuracy among positive samples.
AUC: Area under the ROC curve.
In addition, the Pearson correlation coefficient was used to estimate feature
importance. Pearson correlation coefficient is defined as:
672 B. Zhang

Cov (x, y) represents the covariance between variables x and y. Variances of the
variable x and y are indicated by and , respectively.

3 Result Analysis

3.1 Feature Analysis

Data Visualization. With data visualization, this paper examined the credibility of
the dataset. Fig. 3, and Fig. 4 gives some of the visualization results. Fig. 5 shows the
linear correlation thermogram as a pre-analysis of the data.
Fig. 3 shows the visualization of the variable "Sleep Time" as an example of
analysis. The dataset is evenly and adequately sampled for all age groups. The rate of
heart disease increases nearly linearly with age, thus inferring a positive correlation
between age and prevalence. The same conclusion can also be deduced from the
results shown in Fig.5. The correlation coefficient between age and prevalence is 0.23,
which is the highest among all variables.

Fig. 3. Count and Proportion of Cardiac Patients by Sleep Time(Photo/Picture credit :

Original )

Similarly, it can be deduced through Fig. 4 “Count and Proportion of Cardiac

Patients by Gender” that the percentage of males having suffered from heart disease is
significantly greater than that of females, which accords with the findings of Walli-
Attaei M et al. [11]. In Fig. 5 “Correlation Heatmap”, the variables that strongly
correlate with the prevalence of heart disease were identified with red circles. A
history of skin cancer, kidney disease, stroke, etc. could increase the chances of heart
disease, the habit of smoking could lead to an increase in heart disease, and obesity is
also a possible factor that increases its prevalence. Meanwhile, the black circles in Fig.
5 mark the variables with strong inner correlations, which lays the foundation for
 Deep Neural Network based Heart Disease Prediction 673

dimensionality reduction methods such as feature screening.

Fig. 4. Count and Proportion of Cardiac Patients by Gender(Photo/Picture credit : Original )

Fig. 5. Correlation Heatmap(Photo/Picture credit : Original )

674 B. Zhang

Feature Selection. Variable screening using a correlation coefficient (Pearson

correlation coefficient) larger than 0.05 resulted in eleven variables with high
correlation rates. Fig. 6 shows the comparative performance of the five machine
learning models before and after feature selection, where AUC is employed as the
evaluation index. Except for the logistic regression model, which has a slight increase
in AUC (marked with a red circle in Fig. 6), the rest of the models appear to have a
different magnitude of decline after feature screening. This might have resulted from
the linear essence of the logistic regression model, corresponding to the Pearson
correlation coefficient. Screening out redundant variables with low linear correlation
could help to improve the predictive performance of linear classification models. In
contrast, the Decision Tree and Boosting integrated learning algorithms with a
nonlinear nature performed poorly, as linear feature screening could have instead
caused their main information to be discarded, resulting in performance degradation.

Fig. 6. Performance Before and After Feature Selection(Photo/Picture credit : Original )

3.2 Initial Results

Fig. 7 presents the assessment results described by the four evaluation indices for
DNN and five machine learning models. All models exhibited high accuracy and
recall_0 rate, with a relatively low recall_1 rate (marked with a black rectangle in
Fig.7). For DNN model, Recall_1 rate was merely 0.06.
Since the total number of category 1 (sick) samples was only 1/10th of the category
0 (healthy) samples, all models tended to give higher weight to the category 0 results.
Thus, the imbalanced nature of the dataset could have affected the Recall_1 rate
tremendously. However, Recall_1 rate is a crucial metric for disease prediction.
Hence, imbalanced strategies must be utilized to enhance the performance of
predictive models.
In Fig. 7, the decision tree model exhibited an exclusively lower AUC value
compared to the others. A special analysis was conducted to gain better insight into
the problem. Fig. 8 shows the cross-validation results for the decision tree model. The
portion of the depth above 10 produced large splits in the training and test set results;
 Deep Neural Network based Heart Disease Prediction 675

presumably the severe overfitting phenomenon was the cause of its low performance.

Fig. 7. Original Model Performance(Photo/Picture credit : Original )

Fig. 8. Cross-validation Curve of Decision Tree Classifier(Photo/Picture credit : Original )

3.3 Imbalance Learning Strategy

This research employed random oversampling, random undersampling, and SMOTE
techniques to handle imbalanced datasets in five machine learning models. For
random oversampling, class 1 samples were increased to five times, hence the ratio of
676 B. Zhang

class 0 to class 1 samples in the training set was about 2:1 after processing. For
random undersampling, the class 0 samples were reduced by 2/3, so the ratio of class
0 to class 1 samples was eventually about 3:1. For the SMOTE strategy, there was no
control over the exact ratio of oversampling and undersampling.
The performance of the processed model is shown in Fig. 9, 10, and 11. All three
strategies reduced the accuracy of the five models. After implementing imbalanced
strategies, the AUC values of four models apart from the decision tree model,
declined. By contrast, Recall_1 rates were improved to varying degrees for all five
models. Notably, Recall_1 rate of the logistic regression model saw the most
significant increase by 82.8% after the SMOTE strategy. It can be concluded that the
imbalanced strategies improved the effectiveness of class 1 predictions by sacrificing
the accuracy of class 0 predictions.

Fig. 9. Accuracy Comparison Before and After Imbalance Learning(Photo/Picture credit :

Original )
 Deep Neural Network based Heart Disease Prediction 677

Fig. 10. AUC Comparison Before and After Imbalance Learning(Photo/Picture credit :
Original )

Fig. 11. Recall_1 Comparison Before and After Imbalance Learning(Photo/Picture credit :
Original )
678 B. Zhang

Besides, this study employed imbalanced strategies for DNN model and compared
the results with an integrated learning strategy. For DNN model, two methods were
utilized. On the one hand, F1_Score, AUC, and Recall metrics were monitored
through training. On the other hand, the decision threshold was reduced from 0.5 to
0.1 to increase the weight of class 1 samples. Fig. 12 shows the results of the
integrated learning strategy and DNN model after imbalanced strategies. Two models
were compared under each evaluation metric. Fig. 12 shows that DNN resulted in
slightly higher evaluation metrics than the integrated learning strategy except for
Recall_1 and Precision_0. The DNN model achieved an accuracy of 0.76, an AUC
value of 0.84, and a Recall_1 rate of 0.77 after applying the imbalanced strategies,
owning the best overall performance. Whereas machine learning models such as
logistic regression models performed worse in terms of accuracy (see Fig. 9), and tree
models generally had poor Recall_1 performance (see Fig. 11). It is hypothesized that
linear models are prone to overfitting class 1 samples after imbalanced strategies, and
tree models are consistently more prone to overfitting class 0 samples. Both have
shortcomings in some aspects.

Fig. 12. Performance Comparison Between EnsembleClassifier and DNN After Imbalance
Learning(Photo/Picture credit : Original )

4 Conclusion

In this article, heart disease prediction was performed by deep neural network DNN
using the dataset from Kaggle. The results were compared with five machine learning
models. The DNN eventually reached a prediction accuracy of 0.76, an AUC of 0.84,
and a Recall_1 rate of 0.77 utilizing feature filtering and imbalanced strategies. The
Recall_1 rate had been significantly improved compared to the initial value (0.06),
achieving an ideal predictive result for heart disease. In contrast, machine learning
models behaved lower individual evaluation metrics due to overfitting problems (e.g.,
logistic regression models tended to overfit class 1 samples after applying imbalanced
 Deep Neural Network based Heart Disease Prediction 679

strategies, and tree models consistently tended to overfit class 0 samples). The results
of this paper can assist people in detecting heart disease. However, due to the black-
box characteristics of machine learning and deep neural network algorithms, this
paper did not conduct an in-depth theoretical study. The enhancement introduced by
the imbalanced strategies can be systematically and theoretically analyzed in the
future.

Reference
1. Lindstrom M., DeCleene N., Dorsey H., Fuster V., Johnson C.O., LeGrand K.E., Mensah
G.A., Razo C., Stark B., Varieur Turco J., Roth G.A.: Global Burden of Cardiovascular
Diseases and Risks Collaboration, 1990-2021. J Am Coll Cardiol. (2022).
2. Kumar Y., Koul A., Singla R., Ijaz M.F.: Artificial intelligence in disease diagnosis: a
systematic literature review, synthesizing framework and future research agenda. J
Ambient Intell Humaniz Comput. (2023).
3. Chandrasekhar N., Peddakrishna S.: Enhancing Heart Disease Prediction Accuracy
through Machine Learning Techniques and Optimization. Processes. (2023).
4. Somani S., Russak A.J., Richter F., Zhao S., Vaid A., Chaudhry F., De Freitas J.K., Naik
N., Miotto R., Nadkarni G.N., Narula J., Argulian E., Glicksberg B.S.: Deep learning and
the electrocardiogram: review of the current state-of-the-art. Europace. (2021).
5. Ayano Y.M., Schwenker F., Dufera B.D., Debelee T.G.: Interpretable Machine Learning
Techniques in ECG-Based Heart Disease Classification: A Systematic Review.
Diagnostics (Basel). (2022).
6. Eman H.A., Rabie A.: Clinical Applications of Machine Learning in the Diagnosis,
Classification and Prediction of Heart Failure. Journal of Electrical Systems. (2024).
7. Botros J., Mourad-Chehade F., Laplanche D.: CNN and SVM-Based Models for the
Detection of Heart Failure Using Electrocardiogram Signals. Sensors (Basel). (2022).
8. El-Hasnony I.M., Elzeki O.M., Alshehri A., Salem H.: Multi-Label Active Learning-Based
Machine Learning Model for Heart Disease Prediction. Sensors (Basel). (2022).
9. Guo C.Y., Wu M.Y., Cheng H.M.: The Comprehensive Machine Learning Analytics for
Heart Failure. Int J Environ Res Public Health. (2021).
10. Kingma D., Ba J.: Adam: A Method for Stochastic Optimization. Computer Science.
(2014).
11. Walli-Attaei M., Rosengren A., Rangarajan S., Breet Y., Abdul-Razak S., Sharief W.A.,
Alhabib K.F., Avezum A., Chifamba J., Diaz R., Gupta R., Hu B., Iqbal R., Ismail R.,
Kelishadi R., Khatib R., Lang X., Li S., Lopez-Jaramillo P., Mohan V., Oguz A., Palileo-
Villanueva L.M., Poltyn-Zaradna K., R S.P., Pinnaka L.V.M., Serón P., Teo K., Verghese
S.T., Wielgosz A., Yeates K., Yusuf R., Anand S.S., Yusuf S.: Metabolic, behavioural, and
psychosocial risk factors and cardiovascular disease in women compared with men in 21
high-income, middle-income, and low-income countries: an analysis of the PURE study.
Lancet. (2022).
680 B. Zhang

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-
NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/),
which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any
medium or format, as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter's
Creative Commons license, unless indicated otherwise in a credit line to the material. If material
is not included in the chapter's Creative Commons license and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain
permission directly from the copyright holder.