Enhancing coronary artery diseases screening: A comprehensive assessment of machine learning approaches using routine clinical and laboratory data

Introduction: 

Coronary artery disease (CAD) stands among the leading global causes of mortality, underscoring the critical necessity for early detection to facilitate effective treatment. Although Coronary Angiography (CA) serves as the gold standard for diagnosis, its limitations for screening, including side effects and cost, necessitate alternative approaches. This study focuses on the development and comparison of machine learning techniques as substitutes for CA in CAD screening, leveraging routine clinical and laboratory data. 

Material and Methods: 

Various machine learning classification algorithms—decision tree, k-nearest neighbor, artificial neural network, support vector machine, logistic regression, and stacked ensemble learning were employed to differentiate CAD and healthy subjects. Feature selection algorithms, namely LASSO and ReliefF, were utilized to prioritize relevant features. A range of evaluation metrics, including accuracy, precision, sensitivity, specificity, AUC, F1 score, ROC curve, and NPV, were applied. The SHAP technique was employed to elucidate and interpret the artificial neural network model. 

Results: 

The artificial neural network, support vector machine, and stacked ensemble learning models demonstrated excellent results in a 10-fold cross-validation evaluation using features selected by LASSO and ReliefF. With the LASSO feature selection algorithm, these models achieved accuracies of 90.38%, 90.07%, and 90.39%, sensitivities of 94.43%, 93.03%, and 93.96%, and specificities of 80.27%, 82.77%, and 81.52%, respectively. Using ReliefF, the accuracies were 88.79%, 88.77%, and 90.06%, sensitivities were 92.12%, 91.66%, and 93.98%, and specificities were 80.13%, 81.38%, and 80.13%, respectively. The SHAP technique revealed that typical and atypical chest pain, hypertension, diabetes mellitus, T inversion, and age were the most influential features in the neural network model. 

Conclusion: 

The machine learning models developed in this study exhibit high potential for non-invasive screening and diagnosis of CAD in the Z-Alizadeh Sani dataset. However, further studies are essential to validate and apply these models in real-world and clinical settings.