SRD-GDT²: A SMOTE-Regularized Dual-Guided Deep Tabular Framework for Accurate Cardiovascular Disease Prediction

Abstract

Cardiovascular disease is one of the major causes of death in the world, and therefore, effective and dependable prediction models are necessary. The models can assist doctors in making informed choices and allow the identification of heart-related conditions in a timely manner. Machine learning systems have shown promise in this field; however, their effectiveness is frequently constrained by class imbalance, noisy samples, and unreliable probability estimations in real-world medical datasets. The suggested SRD-GDT² is a system for predicting cardiovascular disease utilizing large-scale tabular data. It is an SMOTE-Regularized Dual-Guided Deep Tabular Distillation framework. The suggested method combines thorough data preprocessing, SMOTE-ENN-based imbalance regularization, and a dual-guided learning strategy in which two tree-based models help a deep tabular learner. An agreement-based weighting system is used to stress dependable recommendations, and Youden's index is used to find the best choice threshold that balances sensitivity and specificity. After preprocessing and resampling a balanced dataset of both majority and minority class samples, experiments were done on a cardiovascular dataset. The suggested model did quite well on the independent test set, with an accuracy of 95.63%, an AUROC of 0.9905, an AUPRC of 0.9870, a sensitivity of 93.79%, and a specificity of 97.51%. The low Brier score of 0.0324 shows that the probability estimations are accurate. Decision curve analysis further illustrates that the suggested framework yields a superior overall treatment advantage when compared with the treat-all and treat-none approaches. Overall, the findings indicate that SRD-GDT² is an effective and clinically valuable method for predicting the risk of cardiovascular disease. It addresses issues related to data imbalance and improves learning stability.