Development of a machine-learning model for predicting arterial hypertension based on complete blood count parameters

Abstract

Objective: development and evaluation of a machine learning model for predicting arterial hypertension using demographic characteristics and complete blood count parameters, and to assess the contribution of key predictors using interpretable methods. Materials and methods: we analyzed a dataset of 5,506 patients examined at the University Clinic of the Medical Research and Educational Institute of Lomonosov Moscow State University in 2021–2022 (2,256 men and 3,250 women, mean age 57.23±17.60 years). Arterial hypertension was identified in 3,697 (67%) individuals, while 1,809 (33%) formed the nonhypertensive group. Variables included sex, age, complete blood count parameters, and derived inflammatory indices (NLR, LMR, PLR). The proposed stacking model was compared with Lasso and Ridge regression, logistic regression, random forest, and gradient boosting. Model performance was assessed using the AUC. Results: AUC ranged from 0.75 to 0.77 for classic machine learning models, whereas the stacking model achieved an AUC of 0.99. The difference in AUC between the stacking model and alternative approaches was statistically significant (p<0.001), with a mean improvement of 22.9% compared with logistic regression. The leading markers were age and the erythrocyte indices RDW-SD, MCV, HGB, and MCH. Conclusion: based on internal validation, ensemble machine learning methods incorporating complete blood count parameters may improve the discriminative performance of cardiovascular risk prediction models; however, these performance estimates require confirmation through external validation in an independent cohort. Age and erythrocyte indices (RDW-SD, MCV, HGB, MCH) were the most informative predictors of arterial hypertension and may serve as accessible marker candidates for further research and improved risk stratification.