OBJECTIVE

To demonstrate on the example of several typical anthropometric, hemodynamic and laboratory indicators a common pattern consisting in the difference of their values’ distributions, as well as to show a non-linear nature of multivariate associations between RF and a probability of fatal outcome by comparing prediction accuracy using several types of ANN.

MATERIAL AND METHODS

The number of men equal 13 263 and women equal 5691 were included with observation up to 2017 from the Russian part of the study of lipid clinics in 1975—1982. They were compared with the publicly available data of study of 3606 men and 2199 women from the USA population (The Third National Health and Nutrition Examination Survey — NHANES III), conducted in 1988—1994, with observation up to 2019. Sex, age, blood pressure, heart rate, body mass index, blood lipid level, smoking and education status, presence of arterial hypertension and hypotension were analyzed. The end point was death from all causes. ANN simulation programs were used.

RESULTS

Differences of the empirical distribution of RF indicators values from normal law according to the Kolmogorov—Smirnov test were statistically significant in all cases with a sample size over 2000 (p<0.01). Comparison of probability prediction accuracy of fatal event by RF values using ANN (linear, probabilistic, radial basic function, three-layer and four-layer perceptrons) showed, that error was minimal in all cases for multi-layer perceptrons.

CONCLUSION

The non-linearity of relationships between risk factors and fatal event is demonstrated by the best result of multi-layer perceptrons that are capable to simulate complex non-linear dependencies compared to artificial neural networks of linear type.