Comparative Analysis of Different Transformerbased Models for Cardiovascular Disease Risk Stratification Using Biomarker Data

Cardiovascular disease (CVD) is a major cause of mortality and morbidity in the world. Risk assessment has been enhanced greatly by the incorporation of multivariate strategies and sophisticated computational models. Recent advancements in deep learning, especially Transformer-based models on tabular data, offer an effective model to capture the complex interactions of biomarkers and improve CVD risk stratification. This paper will suggest a predictive model based on statistical and clinical data of various biomarkers. The conventional machine learning (ML) models compared to deep learning models and three Transformer-based architectures carotid Self-Attention and Intersample Attention Transformer (cSAINT), carotid Feature Tokenizer (cFT), and carotid Tab Transformer. Other experiments evaluate how data augmentation, learning behaviour, and recurrent model fusion impact predictive performance. The results of the area-under-the-curve (AUC) indicate that the Transformer-based models perform better than the traditional ML and DL models in the risk prediction of CVD. In particular, cSAINT had AUC of 0.8875, cFT of 0.9489 and carotid Tab Transformer had the highest performance of 0.9728. This evidence proves that carotid Tab Transformer is a successful model that can be used to model both categorical and continuous biomarkers, with a higher predictive performance and good generalization. Altogether, the suggested Transformer-based model offers a clinical-scale, scalable approach to early risk of CVD and individualized patient care.