Comparison of risk models for mortality and cardiovascular events between machine learning and conventional logistic regression analysis

Aims: Non-linear models by machine learning may identify different risk factors with different weighting in comparison to conventional linear models. Methods and results: The analyses were performed in 15,933 patients included in the Shinken Database (SD) 2004–2014 (n = 22,022) for whom baseline data of blood sampling and ultrasound cardiogram and follow-up data at 2 years were available. Using non-linear models with machine learning software, 118 risk factors and their weighting of risk for all-cause mortality, heart failure (HF), acute coronary syndrome (ACS), ischemic stroke (IS), and intracranial hemorrhage (ICH) were identified, where the top two risk factors were albumin/hemoglobin, left ventricular ejection fraction/history of HF, history of ACS/anti-platelet use, history of IS/deceleration time, and history of ICH/warfarin use. The areas under the curve of the developed models for each event were 0.900, 0.912, 0.879, 0.758, and 0.753, respectively. Conclusion: Here, we described our experience with the development of models for predicting cardiovascular prognosis by machine learning. Machine learning could identify risk predicting models with good predictive capability and good discrimination of the risk impact.