To achieve accurate and comprehensive results, it is essential to have a thorough understanding of the dataset being analyzed [8], [20]. This section examines the dataset, focusing on its features, distributions, and variable correlations. The insights gained will enhance model implementation, clarify the relationship between predictive variables and heart disease, and improve the interpretation of PCA results. The dataset includes 1,025 training samples and 303 test samples, which were intentionally separated to avoid data leakage and bias [17]. Notably, no missing value was identified. Since understanding each feature is critical for contextualizing the research, Table 1 provides definitions and types of each variable.

For improved comprehension, Fig. 1 displays the correlation matrix as a heatmap, illustrating the relationships between every pair of variables. This analysis is crucial, as frequently high correlations among predictive variables can indicate dependency and information redundancy, potentially introducing noise and exacerbating the curse of dimensionality. Additionally, if the correlation of any predictive variable with the target is either extremely low or high, these features may pose a challenge since variables with minimal impact or excessive correlation can act as noise or merely replicate the target, leading to unnecessary dimensionality and compromising the model's robustness and performance [21], [22]. Fortunately, the analysis did not reveal significant concerns, allowing for the inclusion of all features in the modeling process.

According to Fig. 1, Chest Pain type (CP) and maximum heart rate achieved by patients (Thalach) exhibit the strongest positive correlations, indicating that as these parameters increase, so does the risk of heart disease. Conversely, exercise-induced angina (Exang) and ST depression induced by exercise (Oldpeak) revealed the most converse correlations, reflecting an inverse relationship with the target. In this context, Fig. 2 provides a more detailed visualization of these, confirming that the four aforementioned attributes are noticeably different between those who do and those who do not suffer from heart diseases, resulting in some precious insights. For instance, although in the healthcare sector, where the stakes are high, analyzing each feature is paramount, these insights can offer valid intuitions about influential attributes [14], [23], assisting decision-makers when time or resources are limited.

Notably, all figures in Section 2, related to EDA, are based on the complete dataset, showing the distribution of train and test data combined. In other words, although the data were initially separated into training and testing sets, they were integrated for exploratory analysis. Afterward, the data were re-separated into their original formats to ensure that model creation, training, and implementation were conducted with distinct datasets, preventing data leakage and bias, thereby enhancing the validity and generalizability of our findings [12].

After re-separating train and test to minimize bias, both datasets were normalized using the mean and standard deviation from only the training data, crucial for enhanced performance of some ML algorithms like K-Nearest Neighbors (KNN) and Logistic Regression, also ensuring that models remain unaware of test data distribution, enhancing robustness and validity [20]. Various conventional ML algorithms and an ANN were then implemented. Model performance and generalization were fine-tuned through grid searching and 5-fold cross-validation, ensuring optimized model configurations and robust outcomes.

PCA, an unsupervised learning technique, was performed to demonstrate how effectively the data dimensions can be reduced while retaining significant information. We progressively reduced the 13 predictive variables to 12, 11, 10, and so forth, down to 2 features while calculating the information loss at each step. This analysis will illustrate the extent to which the data can be compressed into lower dimensions while being informative. This analysis primarily provides insights into the importance of dimensionality reduction, especially in high-dimensional data scenarios.

After applying and optimizing models, they were evaluated on the test unseen dataset using four metrics: Accuracy, precision, recall, and F1 score. The parameters yielding the reported performances, along with their results, are summarized in Table 2 Most models demonstrated nearly 100% accuracy, highlighting the effectiveness of the employed models on the current dataset.

PCA inevitably results in some degree of information loss due to the dimension reduction and decreased variances. The extent of this loss is influenced by the variance of the features and their correlations. So, Table 3 demonstrates the information loss when reducing data from 13 dimensions to fewer variables. For example, condensing 13 predictive variables into 11 retains approximately 93.84% of the information, but reducing it further results in 90% conserved information. Since effectively summarizing data dimensions is paramount in the health sector, given the high stakes involved, this analysis is immensely beneficial.

In this study, we evaluated model performance using accuracy, precision, recall, and F1 score for two main reasons. First, while accuracy is commonly used, it can be misleading with imbalanced datasets. However, our dataset was relatively balanced in terms of target, reducing this concern. Secondly, accuracy alone does not reveal misclassification types or locations. To address this, we included precision, recall, and F1 score. The recall is crucial as it is sensitive to false negatives, where diseased individuals are misclassified as healthy, potentially leading to severe consequences like death. Thus, the cost of false negatives is higher than that of false positives [24]. Therefore, recall is prioritized over precision in our study. As shown in Table 2, Logistic Regression, despite having the lowest overall performance, achieved a better recall compared to other criteria, which is good. Additionally, the F1 score is valuable and studied as it balances both types of misclassifications by being the harmonic mean of precision and recall.

The study's outcomes are highly beneficial for the health sector. Studying the correlation matrix, feature distribution, and analysis of important features provides valuable insights, making it easier for healthcare professionals to make informed suggestions. For instance, Fig. 1 and Fig. 2 indicate that doctors should advise patients to visit them if they experience intense CP or a rapidly increasing heartbeat during physical activities. In other words, according to the data, they are influential factors to consider.

Furthermore, analyzing all features in practical applications can often be costly and time-consuming. Therefore, focusing on the most important features is crucial, and EDA plays a pivotal role. Therefore, we concentrate on high correlations between predictive variables and the target and also among predictive variables themselves to evaluate feature dependency and avoid the curse of dimensionality. PCA was also shown to be a viable solution to address these concerns by summarizing data dimensions, despite information loss. Since we achieved 100% without dimension reduction in this study, PCA only revealed some theoretical insights. Nonetheless, insights can be highly beneficial and practical in other applications.

Moreover, the study's impressive results, with nearly 100% accurate predictions across all criteria and no overfitting or bias, are noteworthy. Despite the relatively small sample size, 1025 for training and 303 for testing, the findings have practical implications for reducing mortality rates. Additionally, the features proved effective for heart disease prediction, yielding excellent outcomes in both simple models like KNN and DT, and complex models like SVC and ANN. The only consideration is to utilize a considerably larger sample size, ideally in millions, and also more features, as they give rise to enhanced generalization and robustness. Since almost all models performed exceptionally well, they allow doctors and health organizations to choose different models based on their tastes and preferences. For example, some doctors may prefer DT to intricate models like ANN or SVC due to their transparency and ease of understanding, while some others may be keen on ANN, knowing that they are more complex, and their performance will not be saturated with large datasets.

For this dataset, code execution on a Kaggle repository achieved 100% accuracy with RFC and 84% accuracy with Logistic Regression, respectively, validating a part of the present work [17]. However, there are similar studies on heart disease prediction. For example, another study used a data science and ML approach, applying various ML algorithms from the Scikit learn library to predict heart disease. In that study, Logistic Regression achieved the best performance among conventional ML models, with 86.49 % accuracy, 91% recall, 82% precision, and an 86% F1 score [25]. The fact that Logistic Regression performed best in that study, while it underperformed considerably in the current study, suggests that all models are worth exploring. Yet, the model that demonstrates the best performance and generalization on unseen test data should be used for practical applications [8], [25].

The most similar and comprehensive study [14] also focused on accurately predicting cardiovascular disease using various ML techniques. This study achieved 100% accuracy and 100% sensitivity with models such as KNN, DT, and Random Forests. The fact that a relatively simple model like KNN performed as well as the more complex Random Forest underscores the importance of examining a range of models. Additionally, while we examined feature importance using a correlation matrix, this study explored feature importance derived from ML models. Comparing the key features identified by different methods can provide valuable insights. For instance, our study found “Cp” and “Thalach” to have the strongest positive correlations with the disease based on the correlation matrix. Conversely, [14] shows slightly different results. While DT, RF, and Logistic Regression all identify “Cp” as the most influential feature, the AdaBoost classifier highlights “Chol” as the most important. Furthermore, although we identified “Thalach” as the second most influential feature, this was only validated by the AdaBoost classifier. In contrast, DT, RF, and Logistic Regression pointed to other features as being more influential.

I. To create a generalized model for a vast population, it is crucial to study large datasets with millions of observations consisting of more attributes and potentially various nationalities for superior analysis.

II. Exploring the curse of dimensionality and dimension reduction methods, especially in the health sector, where frequent high-dimensional data needs summarization for human interpretation.

III. Integrating diverse data types, such as tabular data, CT/MRI images, heartbeat alterations, and brain signals.

This combination can be groundbreaking research as it enhances prediction accuracy and sheds light on complex patterns and hidden correlations with better EDA.