Determination of the risk of developing diabetes mellitus based on the patterns of symbolic dynamics of the amplitude-time heart rate variability

Abstract

With the sharp rise in the incidence of diabetes worldwide, the screening diagnosis of this disease is one of the most pressing public health concerns. A model that allows differentiating patients with and without diabetes mellitus is proposed in this study. The target group consisted of diabetics of different ages and sex, and the control group included conditionally healthy volunteers and patients suffering from somatoform autonomic dysfunction, ischemic heart disease, nodular goitre and peptic ulcer disease. Lead 1 of the standard electrocardiogram (ECG) was used as the primary signal for analysis. The ECG derivative was the amplitude-time variability of the sequence of R-peaks with the subsequent calculation of the features of symbolic dynamics. They became the input data for the gradient boosting models, which are the best for solving of this class problems. Six different symbolic dynamics techniques have been applied. The study has shown that the amplitude-time variability characteristics prevail over the time variability ones. The LightGBM model has turned out to be the best, showing sensitivity at 0.90 and specificity at 0.94, which was 2-3% better than XGBoost models. The specificity of the model for public databases was at the level of 0.74-0.91, which is an acceptable value for non-invasive diseases screening methods in medical practice.