Prototyping neural networks to evaluate the risk of adverse cardiovascular outcomes in the population

Aim. To develop a neural network basis for the design of artificial intelligence software to predict adverse cardiovascular outcomes in the population.Materials and Methods. Neural networks were designed using the database of 1,525 participants of PURE (Prospective Urban Rural Epidemiology Study), an international, multi-center, prospective study investigating disease risk factors in the urban and rural areas. As this study is still ongoing, we analysed only baseline data, therefore switching prognosis and diagnosis task. Because of its leading prevalence among other cardiovascular diseases, arterial hypertension was selected as an adverse outcome. Neural networks were designed employing STATISTICA Automated Neural Networks (SANN) software, manually selected, cross-validated, and transferred to the original graphical user interface software.Results. Input risk factors were gender, age, place of residence, concomitant diseases (i.e., coronary artery disease, chronic heart failure, diabetes mellitus, chronic obstructive pulmonary disease, and asthma), active or passive smoking, regular use of medications, family history of arterial hypertension, coronary artery disease or stroke, heart rate, body mass index, fasting blood glucose and cholesterol, high- and low-density lipoprotein cholesterol, and serum creatinine levels. Our neural networks showed a moderate efficacy in the virtual diagnostics of arterial hypertension (84.5%, or 1,289 successfully predicted outcomes out of 1,525, area under the ROC curve = 0.88), with almost equal sensitivity (83.6%) and specificity (85.3%), and were successfully integrated into graphical user interface that is necessary for the development of the commercial prognostication software. Cross-validation of this neural network on bootstrapped samples of virtual patients demonstrated sensitivity of 82.7 – 84.7%, specificity of 84.5 – 87.3%, and area under the ROC curve of 0.88 – 0.89.Conclusion. The artificial intelligence prognostication software to predict adverse cardiovascular outcomes in the population can be developed by a combination of automated neural network generation and analysis followed by manual selection, cross-validation, and integration into graphical user interface.