Advanced detection of coronary artery disease using a GAN-transformer model on plasma cytokine profiles

Coronary Artery Disease (CAD) has been a major cause of morbidity and mortality around the world, and the existing methods of diagnosis can be found to detect the disease at an advanced stage. Plasma cytokines, as the main mediational factors of inflammation, are promising as a non-invasive technique of early diagnosis but are extremely dimensional and imbalanced in classes, which significantly complicates their analysis. Our hypothesis in this paper is that a hybrid deep learning architecture can be used to combine a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) to augment synthetic data and a Transformer-based predictor to predict CAD based on plasma cytokine samples. The analysis of a sample of 1,040 patients (390 CAD patients and 650 controls) employed a 45-plex ELISA assay, from which 36 biologically relevant cytokines were selected. To achieve methodological rigor, preprocessing, feature selection, and synthetic data generation steps were carried out in cross-validation folds to avoid data leakage. The analysis of the framework was done with stratified 10-fold cross-validation and an independent external validation cohort. The model achieved an average accuracy of 0.93 ± 0.02 (95% CI: 0.91–0.95), precision of 0.88 ± 0.02, recall of 0.99 ± 0.01, F1-score of 0.93 ± 0.01, and AUC-ROC of 0.98 ± 0.01. External validation gave an accuracy of 97% and an AUC-ROC of 0.996. Jensen-Shannon divergence supported the synthetic data analysis, focusing on MAE, RMSE, and the high fidelity between the actual and generated distributions of cytokines. Such comparative analysis resulted in substantial improvement compared to baseline models, such as logistic regression and traditional machine learning methods. Decision Curve Analysis was more clinically useful, which meant that it offered better net benefit at threshold probabilities. Nevertheless, there are such limitations as lack of diversity among the cohorts, absence of large comorbid populations (e.g., diabetics), and the use of cytokine-only characteristics without the inclusion of clinical variables. In these studies, the results suggest that the proposed GAN-Transformer model is a strong, interpretable, and clinically meaningful tool for early CAD detection using non-invasive biomarkers.