Evaluation of Feature Attributes for the Prediction of Heart Disease Using Machine Learning

Organization of paper

Section 1 describes the introduction of heart diseases and their death ratios globally. It also describes how machine-learning models can address this with minimum time. Section 2 describes the related work on detecting heart diseases using machine learning. Section 3 describes the proposed methodology for detecting machine learning methods. Finally, conclusions on the proposed work and future directions are provided.

2.1. Details of dataset

The dataset has 13 features, including one feature as a target (which indicates healthy and unhealthy states), and it is extended to every patient. This is a balanced dataset with both categorical and numerical variables, which is ideal for developing predictive models and analyzing heart diseases. When an attribute has less than ten different classes, it is considered categorical or nominal. Some of them include sex, type of chest pain, fasting blood glucose level, ECG findings, and many more. The gender attribute is binary, with ‘1’ for male and ‘0’ for female based on the sex attribute. Chest pain type (cp) is categorized into four distinct classes: These subtypes include typical angina, atypical angina, non-anginal chest pain, and asymptomatic chest pain. Fasting blood sugar (fbs) was also binary, depicting a value of more than 120 mg/dL (1 for true, 0 for false).

Three classes were used to classify the results of the resting electrocardiogram (resting): normal, ST-T wave changes, and definite LVH. Another binary variable, exercise-performed angina (exang), assigns a value of 1 to indicate the presence of chest pain during exertion and a value of 0 to indicate its absence. There were three types of slopes for the peak exercise ST segment: downslope, flat, and upslope. While the fluoroscopy visualization of the number of major vessels (ca) ranges from 0 to 3, the thalassemia attribute (thal) categorizes the heart status as either normal, fixed defect, or reversible defect.

The numerical features return continuous data, which is valuable for fine-grained analysis, as opposed to nominal features. These variables included age, resting systolic blood pressure in millimeters of mercury (trestbps), serum cholesterol in milligrams per deciliter (chol), maximum achieved heart rate (thalach), and ST segment depression compared to rest (oldpeak). They record specific clinical parameters that are crucial for assessing the cardiovascular function.

It is also critical to recognize that the target attribute has been transformed into a binary variable from its original quantification into five classes that represent varying degrees of heart disease risk. This simplifies the problem by transforming it into a classification problem, with the goal of ascertaining whether a person has heart disease [ Table 1].

In order to increase number of patients in the dataset combining two datasets namely

The two datasets shared almost the same features, but different names were provided to the features. Table 2 shows naming conventions; the thallium feature is not present in 2 cardiovascular diseases, hence leaving these features and considering the rest of the features. The statistical data for each attribute, including the minimum, maximum, mean, standard deviation, 25%, 50%, and 75%, are displayed in Table 3. As a result, various machine learning models have been trained using a combined dataset to identify the classifier that is most effective in detecting CVD.

2.2. Pre-processing of dataset

If no value is assigned to a particular variable in an observation, it is referred to as missing or incomplete. These missing values can originate from different circumstances, such as when the respondent fails to answer some questions, failure of the sensor, data loss during transfer, disruption in the network connection, or a mathematical operation, such as division by zero. In datasets [ Figure 2], missing values can be indicated by spaces, hyphens, or other marks that differentiate them from other regular values.

Figure 2. Visualization of features in dataset.

The missing values may or may not affect the statistical validity of the outcome. However, the outcomes may lack robustness or accuracy due to omitted information, even though analysis can move forward with incomplete data. Even if each individual variable has only a small percentage of missing data, the cumulative amount across the dataset may be significant, thereby influencing the analysis results. It is worthwhile to replace the observations rather than deleting them because observations with missing values can be quite informative. The strategies include:

2.3. Classification by machine learning models

Earlier studies used multiple supervised machine learning models on a single dataset to predict CVD. Six machine learning models are discussed in this section to predict patients who may be at risk of CVD.

Regression models estimate the target variables on a continuum rather than grouping the outcomes into categories. Linear regression is the basic form of regression analysis, which assumes linearity and attempts to minimize the error between the actual and predicted values. Other types include polynomial and ridge regression, which are for more complex relationships, or when regularization is required.^22,23

Naïve Bayes is a probabilistic classifier that relies on the assumption of feature independence and is derived from the base Bayes formula. This algorithm is computationally efficient and is widely applied to text categorization, spam detection, and sentiment analysis. However, it is efficient and can work well in various real-life situations, particularly when dealing with big data. Based on the input features, it assigns an instance to the class with the highest posterior probability.²⁴

Random Forest is another type of meta-cascade that is composed of many decision trees, where the bagging technique is adopted for increased precision and stability. A subset of the data is used to train each tree, and the final decision (class for classification or mean for regression) is made by considering the results of all the trees. This is less prone to overfitting than individual decision trees, and is flexible for tasks involving structured data.²⁵

Logistic regression is a classification algorithm that predicts probabilities by applying a sigmoid function, which is appropriate when the output is binary. Although closely related to linear regression, it transforms the results into the range of [0, 1]. It has been applied in disease prediction, email classification, and customer churn analysis because of its simplicity and interpretability.²⁷

The gradient-boosting framework XGBoost is popular and often wins machine-learning competitions because it is built for high performance. It iteratively constructs decision trees, gradually corrects the mistakes made in prior steps, and employs shrinkage to curb overlearning. It performs well when dealing with missing values and large structures.¹⁴ Modifications were performed to avoid under-and overfitting.

These modifications in the model help to fight during imbalance dataset and avoids overfitting.

LightGBM is a gradient boosting framework that supports multiclass classification and regression problems and is based on a speed-optimized decision tree. It contains a histogram-based algorithm to build a decision tree for the learning process and manages the memory well. Its approach, such as leaf-wise tree growth, makes it faster than XGBoost for a variety of tasks including recommendations, clicks, and ranking.²⁶

2.4. Features evaluation

To select important features for improving classification that are correlation-based (with two 0.75 and 0.5), chi-square feature selection is employed in the current research. Correlation describes the similarity between two components. This method determines which feature is highly efficient in predicting targets individually. The highly correlated features [ Table 4] between each other, have considered two thresholds to collect correlated values are 70% and 50% of target values. These features have low redundancy and are highly relevant to the target classes.²⁸

Another feature selection method called chi-square attribute evaluation is the arrangement of filters in the order of the computed chi-square statics²⁹ of 10 features. It computes the score of all features with the target values and selects the top ten features. The attributes selected from the chi-square [ Table 5] by eliminating the last two features and training with chi-square-suggested features using machine learning algorithms.

3.1. Dataset information

The dataset employed for all cases is initially discussed, and the results for different cases are provided in the following subsections.

The instance of the employed dataset is obtained by combining the two datasets [ Figure 3], namely Cleveland, Hungary, Switzerland, and the VA Long Beach²³ with 303 patient data and Cardiovascular Disease²¹ with 1000 patient data.

Figure 3. Instance of employed dataset with visualization of counts.

The dataset information [ Figure 4], which includes the total number of entries, number of features recorded per patient, data type, and count of NaN values (black, NA data, or missing values).

Figure 4. Information of dataset.

The distribution of healthy and unhealthy patients [ Figure 5] indicates that the dataset is imbalanced in nature. To address this, a few modifications are made to the XGboost algorithm, as stated in Section 3.3. To visualize missing values, a bar chart is plotted using msno, which provides data points for every column if missing values are present in any column gaps present in the bar charts. There is no discontinuity in the graph [ Figure 6], which clearly indicates that there are no missing values in any column, and the 1303 values indicate that all columns have 1303 entries in the dataset. The heat map [ Figure 7] for the dataset is computed, which gives Pearson correlation coefficients and denotes −1,+1, and 0 + 1, indicating that the value is highly correlated, −1 is a negative correlation, and 0 means no correlation between the target and features. The data [ Figure 8] for training and testing values. If 50% is shown, then the dataset is balanced.

Figure 5. Count of healthy and unhealthy CVD patients in Dataset.

Figure 6. Missing values are demonstrated.

Figure 7. Heat map of dataset.

Figure 8. True and false cases allotted for training and testing in dataset.

3.2. Calculation of Performance metrics for Full Feature Attributes

The performance metrics are listed in Table 6, which includes the confusion matrix, accuracy, precision, recall, and F1 score for the different machine learning models employed. For the analysis in this section, we consider a full set of features. A full set of features means that all 13 features are considered for predicting the health of a patient. Eight machine-learning models have been considered for prediction. To determine the model’s performance, we considered metrics such as the accuracy, precision, recall, F1 score, and confusion matrix. Of these machine learning models, random forests outperformed in the detection of diseases. It achieved a high accuracy of 97.77 and rest metrics such as precision, recall, F1 score of 98.

In addition to the accuracy, the model achieved good values in other metrics, which shows that the model is robust to real-world data. In the confusion matrix, the predictions of True Positive (TP) and true negative (TN) were quite good; there were fewer than 12 and 9 patients that were misclassified as false positive and false negative, respectively. This indicates that the misclassification from the model was comparatively less. A graphical representation of the performance metrics considering the full feature set. Figure 9 shows the performance metrics in a visual chart format. This provides a clear view of which model provides greater accuracy in detecting the heart condition of a patient.

Figure 9. Graphical representation of performance metrics for full feature sets.

3.3. Calculation of Performance metrics for correlation 0.75

Performance metrics [ Table 7] by considering correlation of 0.75 and 0.50 and feature names are chest pain type, exercise angin and st_slope. It also includes the confusion matrix, accuracy, precision, recall, and F1 score for the different machine-learning models employed. The analysis in this section did not consider the full set of features. Instead of all features, 0.75 correlation features are selected. This means that out of 13 features, three correlated features were selected to predict the health of the heart. We considered eight machine learning models for prediction. To determine the model performance, we considered metrics such as the accuracy, precision, recall, F1 score, and confusion matrix. Of these machine learning models, LightGBM outperformed the others in the detection of diseases. It achieved a high accuracy of 77 and metrics, such as precision, recall, and F1 scores of 76, 77, and 76, respectively. As the number of features decreases, it becomes difficult for the model to learn. The accuracy of the model is lower because three features are considered to train the model. Similarly, in accordance with the accuracy, other metric values are also lower. The confusion matrix shows TP and TN values of 186 and 116, respectively, which are good, but not excellent, to adopt models for real-world patient analysis. False positives and negatives are 54 and 35, respectively, which give incorrect predictions. Due to the wrong prediction of healthy as unhealthy and unhealthy patients as healthy lands, the entire family is at risk. Unhealthy patients are left untreated, and healthy patients undergo unnecessary treatment and financial overhead. A graphical representation of the performance metrics considering the full feature set. The performance metrics [ Figure 10], making them easier to interpret than tables. This helps to highlight which model achieves higher accuracy in identifying heart conditions in patients.

Figure 10. Performance metrics by considering feature set of 0.75 correlated values.

3.4. Calculation of Performance metrics for correlation 0.5

Performance metrics [ Table 8] by considering correlation of 0.75 and 0.50 and feature names are sex, chest pain, fasting blood sugar, exercise_angin and st_slope. The analysis in this section did not consider the full set of features. Instead of all the features, with 0. Five correlation features are selected. This means that out of 13 features, three correlated features were selected to predict the health of the heart. Eight machine-learning models have been considered for prediction. To determine the model’s performance, we considered metrics such as the accuracy, precision, recall, F1 score, and confusion matrix. Of these machine learning models, random forest outperformed in the detection of diseases. It achieved a high accuracy of 80.37, and other metrics, such as precision, recall, and F1 score, were 80. As the number of features decreases, it becomes difficult for the model to learn. The accuracy of the model is lower because three features are considered to train the model. In a similar fashion, in accordance with the accuracy, other metric values are also less. The confusion matrix shows TP and TN values of 103 and 76, respectively, which are good but not excellent for the adoption of models for real-world patient analysis. False positives and negatives are 103 and 76, respectively, which give incorrect predictions. Owing to incorrect predictions, there will be unnecessary overhead for the respective families. A graphical representation of the performance metrics considering the full feature set [ Figure 11] provides a good overview of an eagle’s eye view by observing line graphs.

Figure 11. Performance metrics for 0.50 correlated values.

3.5. Calculation of Performance metrics for chi-square

The performance metrics [ Table 9] using the chi-square feature matrix, and the feature names are age, chest pain, resting_bp, fasting blood sugar, max_heart_beat, exercise_angin, oldpeak, st_slope, and num_vessls_fluro. The analysis in this section did not consider the full set of features. Instead of all feature sets, we selected features using chi-square. This means that out of 13 features, nine features were selected to predict the health of the heart. Eight machine-learning models have been considered for prediction. To determine the model performance, metrics such as accuracy, precision, recall, F1 score, and confusion matrix were used. Of these machine learning models, random forest outperformed in the detection of diseases. It achieved a high accuracy of 98 and other metrics, such as precision, recall, and F1 score of 98. The selected features are robust and efficient, facilitating efficient detection. In addition to accuracy, precision, recall, f1 score and confusion matrix were used to measure the efficiency of the model. The confusion matrix shows TP and TN values of 495 and 396, respectively, which are good, but not excellent, to adopt models for real-world patient analysis. False positives and negatives are 13 and 8, respectively, which give incorrect predictions. There are a few numbers of wrong predictions, which help in providing excellent results in detecting the health of a patient’s heart. A graphical representation [ Figure 12] of the performance metrics considering the full feature set. A clearer view can be seen in the chart compared to the table, which makes it easier to understand which model’s performance is better than the other seven machine-learning models.

Figure 12. Performance analysis for chi-square selected feature set.

In printed volumes, illustrations are generally black and white (halftones), and only in exceptional cases, and if the author is prepared to cover the extra cost for color reproduction, are color pictures accepted. If color illustrations are necessary, please send color-separated files if possible. Color pictures are welcomed in the electronic version at no additional cost. The current study utilized two datasets, which were combined by leaving the data from the Cleveland dataset and combined with the CVD dataset to make 1303 patients to predict CVD risk. With all feature sets, the maximum accuracy achieved was 97.77% the random forest algorithm. Not only is accuracy considered, other performance metrics such as f1 score (98%), recall (98%), precision (98%) models are considered as models that provide robustness in detection. The feature selection processes considered were correlation-based and Chi-square-based feature selection. Results analysis showed that the chi-square selected features list achieved a high accuracy, precision, f1 score and recall. Highest accuracy achieved compared to full feature set the chi-square feature selection method was 98%.