The main ways to assess cardiovascular risk are SCORE, SCORE-2, Framingham scales, etc. However, these scales have several statistical drawbacks that may reduce their predictive value. Therefore, many researchers are encouraged to use machine learning technologies to predict cardiovascular risk.
OBJECTIVE
To compare the predictive value of the SCORE, Framingham and one of the machine learning methods based on the «INTEREPID» study.
MATERIAL AND METHODS
The study was conducted based on the data from the international prospective study «INTEREPID» conducted in 2011—2016 (n=2391). Endpoints: cases of coronary heart disease and acute cerebrovascular accident for Samara (n=253; 24.1%) and for the Kyrgyz Republic (n=280; 20.9%), total n=533 (22.5%). ExtraTreesClassifier was used as a machine learning algorithm for the Samara cohort, GradientBoostingClassifier for the Kyrgyz cohort and the entire «INTEREPID» cohort. The SCORE and the Framingham scales were used for comparison.
RESULTS
For the Samara cohort, the best AUC indicator is ExtraTreesClassifier — 0.609; for the Kyrgyz cohort — Framingham scale with an AUC of 0.828; and for the entire cohort, GradientBoostingClassifier with an AUC of 0.766.
CONCLUSION
The results of this study indicate that the quality of determination by the level of cardiovascular risk is better in most cases when using machine learning algorithms; however, in some cases, conflicting data were obtained. In both cases, attention should be paid to the quality of the sample based on which the mathematical model was built and the validation methods.