Improved C-Index and Interpretability with KAN as Compared to COX: Application to Risk Prediction of Major Adverse Cardiovascular Events from CT Calcium Score

Accurate prediction of major adverse cardiovascular events (MACE) remains a significant clinical challenge due to the complexity of cardiovascular risk factors. Epicardial adipose tissue (EAT) has emerged as a promising imaging biomarker for MACE through its association with coronary inflammation. However, existing predictive models predominantly utilize simplistic EAT features, such as volume and mean Hounsfield Units, which limit their effectiveness. In this study, we analyzed 2,316 patients who underwent CT calcium score (CTCS) imaging and extracted a comprehensive set of quantitative EAT features, termed fat-omics. We assessed the predictive performance of traditional Cox Proportional Hazards (CoxPH) models and deep survival analysis methods. While the CoxPH model is widely used, it struggles to capture nonlinear relationships. DeepSurv, employing multi-layer perceptrons, introduces nonlinearity but sacrifices interpretability - a critical factor in clinical decision-making. To address these limitations, we employed CoxKAN, which integrates Kolmogorov-Arnold Networks with survival analysis, enhancing both model expressiveness and interpretability through learnable nonlinear functions. Our results indicate that CoxKAN significantly outperforms both CoxPH and DeepSurv in MACE prediction (C-index: 0.7230 vs. 0.7052 vs 0.6823; 2-years AUC: 0.7337 vs. 0.7238 vs 0.6991, respectively) and offers improved interpretability over purely black-box approaches. By combining advanced fat-omics features with a more transparent deep survival model, CoxKAN may facilitate accurate and explainable MACE risk stratification in clinical practice.