Model for predicting the severity and outcome of COVID-19 in patients with diabetes mellitus and obesity created based on artificial intelligence methods

Background. According to WHO, type 2 diabetes mellitus and obesity are non-communicable epidemics of the 21st century. At the end of 2020, the world community expected the development of the COVID-19 pandemic, which became an epidemic of infectious genesis. Currently, there is a need to identify risk factors for severe course and high mortality in one of the most vulnerable groups of the population with metabolic disorders (DM and obesity) in order to improve the prognosis of COVID-19.

Objective. Development of the model for predicting the severity and outcome of COVID-19 in patients with diabetes mellitus and obesity to optimize diagnostic/treatment tactics.

Methods. The study was conducted in two directions: a retrospective analysis and a prospective part. The retrospective analysis included 645 patients with COVID-19 (58.6% women, 41.4% men). The mean age of patients was 63.6 ± 0.9 years. Diabetes mellitus and obesity occurred in 48.8% (n = 315) and 45.5% (n = 290) of cases, respectively. To identify the features of the clinical course and predictors of an unfavorable prognosis of COVID-19, all patients were divided into 2 groups: group 1 – with recovery (n = 443), group 2 – with an unfavorable outcome (n = 202), between which a comparative analysis was carried out. To develop a prediction model using artificial intelligence (AI) methods, several types of machine learning (ML) algorithms were implemented, among which logistic regression was selected. Regression analysis allowed to classify patients into 2 groups. Class 0 – no risk of adverse outcome and class 1 – high risk of adverse outcome. The prospective part included 130 patients, who formed the validation sample for our study. All patients were distributed according to the severity of COVID-19; demographic, clinical, anamnestic, laboratory and instrumental data from archival medical records were assessed.

Results. The analysis was carried out using 4 main ML algorithms: logistic regression, random forest, support vector machine, gradient boosting. The logistic regression model was chosen for further use in this task, since it showed the highest results for all key metrics, including accuracy, recall and ROC-AUC. The set of basic parameters were represented by the following features: gender, age, day from the onset of the disease, anthropometric data (body weight, height) based on which the body mass index (BMI) is automatically calculated, concomitant diseases (DM, coronary artery disease, arterial hypertension, chronic kidney disease, chronic obstructive pulmonary disease, bronchial asthma), medications taken (insulin therapy, oral hypoglycemic agents, glucocorticosteroids) and laboratory parameters (glucose, creatinine, cholesterol, uric acid, interleukin-6, leukocytes, D-dimmer, total protein, creatinine, urea, aspartate aminotransferase, prothrombin index, C-reactive protein levels) and respiratory function parameters (chest MSCT, respiratory rate, SpO2). Based on the introduction of these data, the model predicted the probable outcome (favorable/unfavorable). After validation in the prospective part of the study, our AI model predicted the risk of an unfavorable outcome with a probability of 96%.

Conclusion. The developed prognostic model allows to prevent the unfavorable course of COVID-19 by timely assessing the severity of the condition and optimizing treatment tactics in the most vulnerable group of patients with metabolic disorders.