A ROBUST AND PREDICTABLE MACHINE LEARNING FRAMEWORK FOR CARDIOVASCULAR DISEASE RISK STRATIFICATION AND PROGNOSIS FOR SURVIVAL

Abstract

Cardiovascular disease (CVD) continues to be the most common cause of death in the world, and the unprecedented growth means that reliable and accurate diagnostic techniques are crucial. Existing machine learning studies are mostly in the binary disease classification domain. Nevertheless, such approach offers limited time-to-event sensitivity and little sensitivity to uncertainty in ML models. This restricts their clinical value substantially. We propose a comprehensive, interpretable machine learning framework in this paper for the risk stratification of cardiovascular diseases and hope of survival. It combines both classical survival analysis plus neural survival models, multi-task learning, and Bayesian uncertainty quantification. Right-censored survival aims were coded and validated using Cox Proportional Hazards, Neural Cox, Neural Additive Survival Networks, Multi-task Neural Networks, and Bayesian Neural Survival Models (where routine collection of clinical parameters is used, respectively). Performance of their experimental results indicated improvement in diagnostic performance [AUC = 0.998] and survival discrimination, leading to better survival discrimination, with concordance index value of 0.706. This framework provides for the generation of understandable risk curves, hazard ratios and credible intervals, from which clinically important and decision risks for dangerous cardiovascular outlook can be derived.