Cardiovascular Therapy and Prevention (Russian Federation) · Published 2026-04-29 · DOI 10.15829/1728-8800-2026-4826
Aim. To study the predictive effectiveness of neural network models for 30-year prediction of fatal outcomes using conventional risk factors.Material and methods. A total of 13263 men and 5691 women from the Russian part of the Lipid Clinics study from 1975-1982 were included, with follow-up through 2017. A comparison was made between 3606 men and 2199 women from the US population from the Third National Health and Nutrition Examination Survey (NHANES III) conducted between 1988 and 1994, with follow-up through 2019. The endpoint was all-cause mortality. Sex, age, blood pressure, heart rate, body mass index, blood lipids, smoking status, education status, and the presence of hypertension and hypotension were analyzed. Artificial neural network (ANN) simulators were used to construct multivariate models.Results. For each of the four cohorts (Russian men and women, US men and women), 1000 3-layer perceptron ANNs were trained. Receiver Operator Characteristic (ROC) curves were constructed for the 500 best models, resulting in a set of 500 ROC curves. We calculated the average areas under the ROC curves for the Russian and US cohorts; they were approximately 0,8 and 0,9, respectively. This corresponds roughly to the maximum percentages of correct predictions of 73-77% in Russia and 83-85% in the US. The stability and consistency of the results in the training and test subsamples in both the Russian and US cohorts demonstrate the adequacy of our approach to constructing predictive models.Conclusion. Neural network models for predicting the probability of fatal outcomes based on 30-year prospective follow-up demonstrated high performance as follows: average areas under the ROC curves in the test subsamples were 0,8-0,9.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →