Detection of Heart Disease Using Binary Classification Machine Learning Model

Heart diseases are atherosclerotic Conditions that impact the heart organ, which include ailments like arrhythmias, heart failure, and coronary artery disease [1]. Smoking, high blood pressure, raised cholesterol levels, bad eating choices, inactivity, and obesity all increase the chance of getting particular cardiac diseases. One of the most common cardiac disorders is coronary artery disease, which is identified by the constriction or obstruction of the coronary arteries, chest pain, heart attacks, and strokes that may result from it. Other heart issues include endocarditis, an inflammation of the inner layer of the heart, congestive heart failure, irregular heartbeats, congenital heart problems apparent at birth, and irregular heart rhythms [2]. The phrase "cardiovascular disease" encompasses conditions that affect the heart [3]. These heart-related disorders can be avoided or controlled with the help of medication if the patient adopts a healthy lifestyle.

Modern healthcare has come a long way, for example [2] made a considerable contribution to the global increase in average life expectancy by introducing intelligent algorithms with high accuracy levels in classifying different heart diseases. The complexity of addressing the changing demands of an aging population, as well as increased service demand and rising expenses, provides difficulties for healthcare systems. As a result, the healthcare sector has produced a vast amount of data about individuals, ailments, and diagnoses. Unfortunately, due to insufficient investigation and analysis, the entire potential of this material stays untapped. Medical experts have a fantastic potential to improve patient prognosis thanks to the expanding body of medical data. Healthcare professionals have been utilizing computers more and more in recent years to enhance decision assistance. For difficult jobs like deciphering genomic data and turning medical records into useful knowledge, machine learning has turned out to be a valuable technique in the healthcare industry.

The United States is one such country where heart disease remains the primary cause of death for individuals of all genders and across various racial and ethnic backgrounds. The alarming fact is that every 33 seconds a person dies of cardiovascular disease in the US [4]. Heart disease will have a significant impact because it will be responsible for 20% of all fatalities in 2021, killing about 695,000 people. Heart disease has a large financial impact as well; it costs the US $239.9 billion annually in healthcare costs, prescription drug costs, and the financial fallout from premature deaths [4].

Cardiovascular disease is not only a leading cause of death globally but also a major source of financial strain on healthcare systems. According to the American Heart Association, the expenses associated with treating cardiovascular conditions in the United States could surge to \$1.1 trillion by 2035, driven by rising disease rates [5]. Worldwide, cardiovascular diseases contribute over \$1 trillion annually in both direct medical costs and indirect losses, such as diminished workforce productivity. This economic burden is particularly severe in low- and middle-income nations, which account for approximately 75% of the global cases, thereby worsening existing financial disparities [6].

Table 1 illustrates research regarding heart disease and its impact on men who are African American, American Indian or Alaska Native, Hispanic, and white, this shows the statistics of heart disease on certain races in the United States.

Table 1. The percentages of heart disease deaths among the various ethnic groups (races) in 2021

In terms of annual mortality, cardiovascular diseases (CVDs) are the leading cause of death globally, outpacing all other causes [7]. A startling 31% of all fatalities worldwide in 2015 were attributable to CVDs, which claimed an estimated 17.7 million lives in that year alone. Within this sobering figure, coronary heart disease was responsible for about 7.4 million deaths while stroke claimed 6.7 million. Notably, low- and middle-income countries occupy the largest share in nearly 75 percent of CVD-related mortality. Unbelievably, 82% of the 17 million premature deaths from noncommunicable illnesses that occurred in 2015 occurred in these low-income areas, with cardiovascular diseases (CVDs) accounting for 37% of those deaths [8].

However, most cardiovascular diseases are preventable and many of them are due to Risk factors that can be changed, such as smoking, unhealthy eating habits, and being overweight, inactivity, and excessive alcohol consumption [9]. Early detection and management are essential for those with cardiovascular disease or those who have a high risk (as a result of conditions like hypertension, diabetes, hyperlipidemia, or pre-existing conditions), and they should include the proper counseling and pharmacological interventions.

Diagnosing heart disease accurately and swiftly remains a significant challenge in medicine despite advances in technology. While existing diagnostic tools have limitations, such as precision issues with electrocardiograms and the risks associated with invasive therapies, there is a pressing requirement for a dependable, non-invasive, and efficient diagnostic method that supports early detection. Machine learning algorithms offer promise in utilizing available data, including routine blood tests and patient demographics like age and gender, to predict future cardiac events accurately. However, previous research using other machine learning models has yielded suboptimal results due to limited datasets and lack of cross-validation, relying solely on baseline models. Leveraging the predictive power of ensemble models like random forests could address these limitations and improve diagnostic accuracy.

This study aims to detect heart disease using a predictive binary classification machine learning model. Therefore, the specific objectives of the study include:

1. Compare and contrast various machine learning methods for predictive binary classification to see how well they can detect heart disease.
2. Create predictive binary classification algorithms that can accurately forecast a patient's likelihood of having heart disease.
3. Implementation analysis and evaluation of the results and performance metrics of various predictive binary classification models.

1.1 Problem statement

Several studies have demonstrated the efficacy of machine learning techniques in predicting heart disease, achieving high levels of accuracy and reliability. For instance, models utilizing algorithms such as Support Vector Machines (SVM), Random Forests, and Gradient Boosting have reported accuracy rates exceeding 90%. These results suggest that machine learning can significantly enhance traditional diagnostic methods by providing more precise risk assessments and enabling personalized treatment plans tailored to individual patients.

Despite the promising results from various machine learning applications in heart disease detection, challenges remain in terms of model interpretability and integration into clinical practice. Ensuring that these models are transparent and understandable is crucial for gaining the trust of healthcare providers and patients alike. Therefore, this research aims to develop a binary classification machine learning model specifically designed for heart disease detection, focusing on achieving high predictive performance while maintaining interpretability. By addressing these critical aspects, this study seeks to contribute to the ongoing efforts to improve heart disease diagnosis and management through innovative technological solutions.

Heart disease continues to pose a significant threat to global health, and the integration of machine learning into diagnostic practices offers a viable pathway toward enhanced early detection and intervention strategies. This research endeavors to harness the power of machine learning to develop an effective binary classification model that can aid in identifying individuals at risk for heart disease, ultimately contributing to better health outcomes and reduced healthcare costs.

For this research purpose secondary data was considered, we utilized the dataset "Heart Failure Prediction Dataset" which can be accessed at the Kaggle database for heart disease prediction [15]. This dataset stands out because of its original integration of different datasets that were not previously combined. Using this vast dataset, an array of machine learning models was created and evaluated with the goal of forecasting and preventing heart illnesses.

K-Nearest Neighbors (KNN), Decision Trees (DT), Logistic Regression (LR), Support Vector Machines (SVM), and Random Forests selection for the detection of heart disease in this study are grounded in their theoretical suitability and practical effectiveness for binary classification tasks. KNN operates on a simple principle of classifying data points based on the majority class among their nearest neighbors, making it particularly useful for heart disease prediction where decision boundaries may not be linear. Its non-parametric nature allows it to adapt to various datasets, capturing local patterns effectively. Decision Trees offer significant advantages in interpretability, allowing for easy visualization of decision-making processes, which is crucial in clinical settings. Logistic Regression, being straightforward and computationally efficient, provides quick insights into the relationship between independent variables and the likelihood of heart disease, outputting probabilities that facilitate risk assessment. Support Vector Machines excel in high-dimensional spaces, making them suitable for medical datasets with numerous features; their ability to apply different kernel functions allows them to adapt to various data distributions while maintaining robustness against overfitting. Random Forests leverage multiple decision trees to improve predictive performance and reduce overfitting through an ensemble approach, offering valuable insights into feature importance that can guide clinical decision-making. By integrating these diverse methodologies, this research aims to develop an effective predictive model that identifies individuals at risk for heart disease while supporting clinical decisions through clear insights into risk factors and underlying patterns.

The research model used in the proposed study is illustrated in the subsequent Figure 1. This dataset is of particular interest because it combines five different cardiac datasets and shares eleven common attributes. As a result, it currently stands as the largest dataset that has been found for heart disease research. The five constituent datasets used in creating this dataset are represented in Table 2.

1.png

Figure 1. Proposed modelling strategy

Table 2. The five constituent datasets

3.1 Data description

The combined dataset comprises data from five distinct cardiac datasets and encompasses a set of 11 shared attributes. It includes a total of 12 characteristics, with 11 independent features and 1 target feature serving as the dependent variable, this is represented in Table 3. With a substantial sample size of 918 observations, this dataset holds the potential for predicting and potentially preventing cardiac diseases.

Table 3. The 12 attributes and their description from the datasets

3.2 Data preprocessing and implementation

In the process of the data preparation for model training, it is important to identify the nominal and ordinal categorical variables in the dataset:

1. Ordinal categorical variables—ST_Slope

2. Nominal categorical variables—ChestPainType, RestingECG, ExerciseAngina, Sex

An ordinal categorical variable is a type of categorical variable that has a specific ranking or order among its possible values. These variables lie between categorical and quantitative variables. For example, education level has a distinct order, such as "high school", "college", and "graduate school".

On the other hand, a nominal categorical variable consists of two or more categories without any inherent hierarchy or order [16]. There is no natural ranking between the categories. A binary variable, such as a yes-or-no question, is an example of a nominal categorical variable, with two distinct categories (yes or no) and no inherent ordering.

Since machine learning models can only process numerical features and cannot directly handle categorical attributes, it is necessary to transform ordinal and nominal qualitative variables to numeric representations. This process is known as feature encoding. By encoding these categorical variables into numerical features, the machine learning model can effectively utilize them for training and prediction purposes. For ordinal categorical variables, we use an ordinal encoder method, as shown below: Where, ordinal_variable = ["ST_Slope"] For nominal categorical variables, we use one-hot encoding method, as shown below: Where, nominal_variable = ["Sex", "ChestPainType", "RestingECG", "ExerciseAngina"].

While the Heart Failure Prediction Dataset provides a robust framework for model training, it is obsereved that the dataset primarily combines records from several hospitals in the U.S. and Europe, which may introduce demographic biases. For example, patients from low- and middle-income regions are underrepresented, limiting the model’s generalizability to a global population. Additionally, the dataset, though free of outliers and missing data, could still suffer from a class imbalance, with heart disease cases outnumbering non-cases. This imbalance might inflate performance metrics such as accuracy, which is why the study employed cross-validation and data-balancing techniques to ensure more reliable results across diverse populations.

3.3 Train-test split

The dataset has to be split into the training dataset and the test dataset to run a machine learning process. The test dataset is used to assess how effectively the machine learning model has learned and how correctly it can generate predictions, while the training dataset is needed in order to train the model. A split ratio of 80:20 was employed for the "Heart Failure Prediction Dataset," using an 80-20 divide applying 80% for training the model and 20% for the evaluation. The train-test divide is put into practice as follows:

(1) The feature "HeartDisease," which serves as the output or target variable, is assigned to a variable called "target_variable" for ease of implementation and understanding.

(2) The "target_variable," which contains the "HeartDisease" feature, is then removed from the dataset.

(3) The modified data is assigned to a variable called "x". Subsequently, the dropped "target_variable" is assigned to another variable called "y".

(4) This separation allows for the independent handling of input features which is encoded as "x" and the variable targeted is stored in "y" during the training and test processes of the machine learning model.

(5) TRAIN: x-train (734 records, 18 attributes), y-train (734 records), distribution of target (401 “heart disease” which is 1, 333 “No Heart Disease” which is 0).

(6) TEST: x-test (184 records, 18 attributes), y-test (184 records), distribution of target (107 “heart disease” which is 1, 77 “No Heart Disease” which is 0).

3.4 Split data scaling (Normalization)

To ensure consistency in how machine learning models interpret the features, the dataset underwent a process called normalization [17]. The process of modifying different characteristics in a dataset is known as feature scaling, also known as normalizing. Real-world data often contains features with different size, scope, and unit values. Feature scaling is necessary so that machine learning models can understand these features on a consistent scale.

Normalization, specifically min-max scaling, was applied to the dataset. This scaling technique involves shifting the values within each column so that they have a set range between 0 and 1 as their bounds [18]. By performing normalization, the dataset's features are brought to a common scale, allowing machine learning models to accurately analyze the dataset. This is crucial because features in the dataset may originally have different magnitudes, ranges, and units. The min-max scaling approach used in this study ensures that the values within each column are adjusted to fit within the range of 0 to 1, promoting consistent and standardized interpretation of the data.

3.5 Model building (Baseline models)

The process of building baseline models involves creating simple reference models that serve as a benchmark in a machine learning project. These models are typically straightforward and may not possess strong predictive capabilities. However, including baseline models is crucial for various reasons [19]. Below is the breakdown of the baseline model building:

1. In this study, baseline models were constructed using several algorithms imported from the scikit-learn library. The specific models used include the Random Forest Classifier from the sklearn.ensemble module, SVC (Support Vector Classifier) from the sklearn.svm module, Decision Tree Classifier from the sklearn.tree module, K-Neighbors Classifier from the sklearn.neighbors module, and Logistic Regression from the sklearn.linear_model module.

2. To train the baseline models, the X-train data (input features) and Y-train data (target variable) were fitted into each respective model. This step involves using the training data to teach the models to make predictions based on the input features.

3.6 Model building: Cross validation model

Several methods that can be used in machine learning, and one of them is the cross-validation method that is used to determine how well fitted a model is for a new-coming data. This involves a process of partitioning the data into two or more subsamples or folds and can be defined as below in other words, it can be defined as a process of data partitioning. The different samples of the given data are divided into one part- this is the validation partition, and the other parts of the data samples are used in training of the model. Thus, this procedure is carried out time and time again, such that each of the folds is used as the validation set [20]. In order to increase the reliability and validity of the results, the performance of the model is evaluated at different epochs and the results are averaged out.

Cross-validation solves the difficulty of overfitting where the model is excessively trained on the training data and will not work well when tested on other data. Cross validation can be used to have a better estimate of how well the model works on other unseen validation sets. The sort of cross-validation include k-fold, the leave-one-out, the stratified [21]. The modeling challenge, the data requirements, the size and kind of the data, and other elements all have a role in the method of choosing.

Implementing cross-validation in the context of model creation often entails the following steps:

1. Add cross-validation modules from the Scikit-Learn package, like StratifiedKFold and GridSearchCV from sklearn.model_selection.

2. Import evaluation metrics from the scikit-learn package, such as confusion_matrix, roc_curve, auc from sklearn.metrics, accuracy_score, recall_score, f1_score, and precision_score.

3. Fit the X-train and Y-train data into the cross-validation modules to train the models.

3.7 Comparison between cross-validation and train/test split in machine learning

Train/test split: It involves splitting the input data into two data sets which includes the training data set and the testing data set in a ratio of 70:30 or 80:20. The training data is used to develop the model while the test data is used to evaluate the model developed. The primary disadvantage of this approach is that it has fairly high variance.

1. Train Datasets: The train set consists of observations used to train the model, where each observation has a known dependent variable.

2. Test Data: The test data is a subset of the original data with similar characteristics, and is separate from the data used for training purposes. It is used to check the performance of the trained model by making predictions.

Cross-Validation dataset: Cross-validation decision is used to avoid the drawbacks of the train/test split approach. In this method, data is divided into a number of partitions and each partition is used as the training and testing sets in turns. It is for this reason that assessments are then taken and averaged in order to arrive at a more accurate conclusion. Cross-validation is most advantageous when it is used in the tuning step of a trained model and all features and records are utilized in the training and testing step, which makes it far superior to the train-test split method.

In the process of building a model for a certain task it is recommended to evaluate a performance of a certain type of a machine learning model. Some of these are confusion matrix, recall, precision and accuracy.

3.8 Model building: Cross validation model

As indicated in the evaluation metrics, the effectiveness of the proposed model is determined by the confusion matrix, the recall rate, the precision, and the accuracy. The confusion matrix is an n × n matrix; the value of n being equal to total number of different target classes in the problem; helps in comparing the actual results with the results predicted by the ML model.

The matrix helps in comparing the actual values of the test set with the values that had been forecasted by the classifier [22]. In an effort to eliminate some level of error the aim is to increase True Negatives and True Positives and decrease False Negatives and False Positives. An optimal model will work towards the achievement of this balance.

True Negative (TN) is a state in which the machine learning model accurately predicts that a patient does not have heart disease, and the patient does not.

False Negative (FN) is a case whereby the model returns an indication that the patient has no heart disease while in the real sense he/she has the disease.

True Positive (TP) means that the model has classified a patient as having the heart disease when in fact the patient does have the disease.

False Positive (FP) means that the model produces results that indicate a certain patient have heart disease but in real sense he or she does not.

Accuracy is another measure that measures how correct the model is in the prediction it gives. However, accuracy is sometimes misleading especially when dealing with many class problems or datasets that have imbalanced classes.

Accuracy $=\frac{T P+T N}{T P+F P+T N+F N}$

A measure of how well a model distinguishes between positive cases, precision looks at the ratio of correctly identified positive cases to the total number of positive cases the model predicted, it is determined by the ratio of correctly deemed positives to the total of well classified positives and

Precision $=\frac{T P}{T P+F P}$

Recall quantifies how many accurately identified positive outcomes there are compared to all number of results in the true positive class. It reveals information on how well a model can locate relevant data.

Recall $=\frac{T P}{T P+F N}$

The F1-score is used to evaluate how well a model performs on a particular dataset, commonly in binary classification situations where data is grouped into "negative" and "positive" categories. The F1-score is determined by taking the harmonic mean of the model's precision and recall results, giving a comprehensive assessment of both measures.

F1-Score $=\frac{T P}{T P+1 / 2(F P+F N)}$

The AUC assesses the size of the 2D area under the ROC curve. It gives an indication of how well various threshold settings perform overall. The ROC curve shows how True Positive Rate and False Positive Rate are related, providing a visual representation of how well a classifier performs.

8.png

Figure 8. Feature importance of cross-validated random forest classifier model

5.3 Feature importance of optimal model

The idea of feature importance entails scoring input features to determine their importance in a particular model. The relative importance of each attribute is shown by these scores as illustrated in Figure 8. A high value suggests that a particular feature has a greater impact on the model's predictive capabilities to predict a given variable [23].

The feature importance analysis reveals the top five influential features in the dataset: ST_Slope, Oldpeak, MaxHR, Cholesterol, and Age using the random forest classifier, the model with the highest performance. This means that any increase in these features will increase the probability of the model predicting the presence of heart disease. As a result, it is advised that patients with a likelihood of developing heart diseases regularly monitor these features and take steps to reduce the weight of these features and any possibilities of an increase in weight. By doing so, it lowers the risk of developing heart disease. Even if the age element might be unavoidable, proactive management of the other features might assist minimize the risk of heart disease.

5.4 Implication of study finding

One of the primary benefits of employing machine learning algorithms is their ability to facilitate early detection of heart disease, which is crucial for effective intervention. Early identification allows healthcare providers to implement timely treatment strategies, potentially preventing the progression of the disease and reducing the risk of severe complications or mortality. This tailored approach not only improves patient management but also optimizes resource allocation within healthcare systems, as clinicians can prioritize interventions for those at the highest risk.

The implications of this study underscore the transformative potential of machine learning in early heart disease detection and risk stratification. By harnessing these advanced technologies, clinicians can enhance diagnostic accuracy, tailor treatment strategies, and contribute to more efficient healthcare delivery systems. As such, integrating machine learning into routine clinical practice represents a promising avenue for addressing the global burden of heart disease and improving patient care.

Machine learning models can augment existing Clinical Decision Support Systems (CDSS) by providing real-time data analysis and predictive insights. These systems can assist clinicians in interpreting diagnostic results, thereby reducing the potential for misdiagnosis and enhancing the overall quality of care. By integrating machine learning outputs into CDSS, healthcare professionals can receive recommendations for further testing or intervention based on predictive analytics, streamlining the decision-making process.