Pathophysiology of metabolic syndrome: new possibilities of artificial intelligence

Abstract

Relevance. Compared to laboratory methods of performing biochemical blood tests, artificial intelligence algorithms offer several key advantages. They are comparable to traditional methods in accuracy and efficiency, but significantly superior to them in cost-effectiveness.

The goal of the study is to develop an interpretable model for calculating a patient’s lipid profile using machine learning based on depersonalized medical examination data (age, gender, complete blood count (CBC), total cholesterol, blood glucose).

Materials and methods. The developed algorithm is a set of technological solutions that allows one to calculate the relationship between various biochemical parameters in the human body using advanced deep learning algorithms, such as gradient boosting on decision trees and fully connected neural networks. The database included the results of laboratory parameters of 62,192 patients: cholesterol (C), high-density lipoprotein (HDL), triglycerides (TG), low-density lipoprotein (LDL), CBC, as well as the sex and age of the patients.

Results. The quality of the models was assessed on the test set using metrics such as the coefficient of determination (R2), mean absolute error (MAE), mean absolute percentage error (MAPE). The R2 for the correlation between LDL concentration and follow-up data group was 0.94 with MAE 0.20 and MAPE 0.06. The correlation between HDL and TG with follow-up data was significantly lower (R2 was 0.51 and 0.41, respectively, MAE 0.20 and 0.53, respectively, and MAPE 0.15 and 0.40, respectively).

Conclusion. The development of an artificial intelligence algorithm that can predict LDL, HDL and triglyceride levels based on less expensive laboratory tests such as cholesterol and blood glucose levels have enormous potential to improve access to healthcare. Although high accuracy (94%) has been achieved to date in LDL-C prediction alone, the next important step is to select the processing method and expand the training dataset to more accurately predict HDL and TG levels.