Using machine learning to assess cardiovascular risk in T2DM patients without established cardiovascular disease: The MARK-2 analysis

Abstract Background:Cardiovascular (CV) risk assessment is the cornerstone for choosing the most appropriate antihyperglycemic agent. One of the major obstacles in the use of such scores is that the process is time consuming, which is a deterrent in a busy clinic. The aim of the present analysis was to explore means of reducing the number of inputs to make the scoring faster and more acceptable to physicians.Methods:We used the MARK survey dataset to assess the utility of the QRISK 3 CV risk score in the Indian population (n=1,538); participants were recruited from five zones (east, west, north, south, and central). Machine learning techniques were used to analyze the existing MARK data. The XGBoost algorithm was used to create the model needed to identify the clinical parameters of interest. Dimension reduction techniques were employed to identify the clinical parameters of importance based on the accrued gain values. As the final step, a confusion matrix was created to assess the model’s accuracy and precision.Results:We identified ten parameters of clinical interest that could predict the 10-year CV risk and were comparable with the QRISK 3 score (with 21 parameters). The parameters identified by machine learning (ML) were age, sex, BMI, smoking status, blood pressure, treatment with antihypertensive agents, angina, total cholesterol, the total cholesterol/HDL ratio, and CKD. This dimension reduction was possible with a model accuracy of 95.12%.Conclusion:There were ten routinely assessed parameters identified with machine learning techniques that can assist physicians in determining the CV risk of T2D patients within a short period of time.