Атеросклероз (Oct 2022)
Immunophenotyping of circulating leukocytes as a tool to optimize diagnosis of carotid atherosclerosis using machine learning approach
Abstract
The aim of the present study was to investigate the possibility of using immunophenotyping of circulating lymphocytes, neutrophils and monocytes to diagnose prognostically unfavorable carotid atherosclerosis using machine learning algorithms. Material and methods. A sample of patients aged 40 to 64 years, who underwent duplex scanning of the carotid arteries and the lower limb arteries, served as a source of patients for analysis. Phenotyping and differentiation of subpopulations of lymphocytes, neutrophils, and monocytes were performed by flow cytometry using the “Navios 6/2” device (Beckman Coulter, USA) device. Data were statistically processed using software package SPSS 23 and our own software programs created using main libraries of Python 3.8 programming language (Scikit-learn, Pandas, Numpy, Scipy) and IDE Spyder. Results. Seventy-eight patients, 39 (50.0 %) males and 39 (50.0 %) females, median age 50.0 years, were included in the study. Age over 56 (р = 0,001), elevated low density lipoprotein (LDL) cholesterol (р < 0.001) and uric acid (р = 0.001), as well as immunosuppressive neutrophils level (р = 0.005) were statistically significantly associated with the development of carotid plaque. At the same time, decreased cell populations of proangiogenic neutrophils (р = 0.009), classical monocytes expressing CD36 (р = 0.019), nonclassical monocytes (р = 0.021), intermediate monocytes expressing TLR4 (р = 0.033), and classical monocytes expressing TLR2 (р = 0.044) were statistically significantly associated with an increased chance of carotid plaque. Taking into account the selected parameters, two prognostic models were created. The first model included basic clinical and laboratory parameters (age, LDL cholesterol and uric acid), and the second model included all selected parameters as well as immunological parameters. Inclusion of the identified immunological predictors in the model resulted in a significant increase in all standard quality metrics of the binary classification. Model accuracy increased by 13 % (р = 0.014), sensitivity by 20 % (р = 0.003), specificity by 6 % (р = 0.046), predictive value of a positive result by 9 % (р = 0.037), predictive value of a negative result by 16 % (р = 0.011). According to the ROC analysis, without the inclusion of immunological predictors in the model, the area under the ROC curve (AUC) was 0.765 [0.682; 0.848], the inclusion of immunological predictors resulted in a statistically significant increase in AUC (0.906 [0.854; 0.958], р = 0.041). Conclusions. In patients 40–64 years old without established atherosclerotic cardiovascular disease, inclusion of immunological markers derived from leukocyte immunophenotyping in the model allowed increasing its diagnostic efficacy with regard to the detection of prognostically unfavorable carotid atherosclerosis. Subpopulations of monocytes expressing TLR2, TLR4, and CD36, as well as immunosuppressive and proangiogenic neutrophils, demonstrated diagnostic value.
Keywords
