Comparative Analysis of the Calculation Models for Ischemic Heart Disease Overall Risk in Railroad Workers

Aim. To conduct a comparison of IHD risk prediction with the SCORE, PROCAM andFraminghamscores, as novel developed models of IHD risk in cohort of the railroad workers. Material and methods. Totally, 106 patients included — workers of locomotive crews, who had been under medical observation during 2006-2015 years with regular scheduled physician observations, incl. preflight, in-depth investigation during in-patient cardiological hospitalization. IHD diagnosis was set if documentarily confirmed signs of coronary ischemia of myocard were found, or coronary atherosclerosis (by angiography, MDCT). Individual total vascular risk was estimated by SCORE, PROCAM, Framingham scores and IHD prediction models developed with the principal components approach (PCA), probabilistic neural networks (PNN), Decision Trees based upon a complex of parameters (including body mass, low density lipoproteids, triglycerids, pulse BP, etc.). Quality of the models was assessed via the parameters of sensitivity, specificity, mean absolute prediction bias, square under ROC-curve (AUC). Results. Dispersion to the categories of low, moderate and high by SCORE, PROCAM andFraminghamwas controversial. In IHD group with low risk by these scores, cardiovascular events (myocardial infarction and/or coronary surgery) happened in 51,9%, 5,56% and 42,6%, respectively. AUCs for SCORE, PROCAM andFraminghamin “IHD” subclass were 0,72, 0,65 and 0,69, resp., but for subclass “Myocardial infarction” — just 0,34, 0,42 and 0,32. AUC for PNN on “IHD” was 0,55, “Myocardial infarction” — 0,60. Mean absolute prediction bias for high IHD risk by SCORE, PROCAM and PNN was 0,88, 0,56, 0,46, resp. Conclusion. SCORE, PROCAM andFraminghamscores are inconsistent for the assessed IHD risk category and are not sufficiently precise for prediction of myocardial infarction in cohort of adult men, workers on railroad, that reduces applicability of the scores. Novel developed model of PNN shows the prediction bias lower in models with added parameters as pulse pressure, platelets, triglycerides, myocardial mass index and other).