A machine learning analysis of predictors of future hypertension in a young population.

Abstract

Background: Early diagnosis of hypertension (HT) is crucial for preventing end-organ damage. This study aims to identify the risk factors for future HT in young individuals through the application of machine learning (ML) models.

Methods: The study included individuals aged 18-40 years who had not been diagnosed with HT through ambulatory blood pressure monitoring (ABPM). These participants were monitored for hypertension diagnosis from the date of ABPM application until the date of data collection. Hypertension prediction was carried out using three distinct ML methods: Support Vector Machine, Random Forest, and Least Absolute Shrinkage and Selection Operator. The identification of variables significant for future HT was based on the outcomes of these models.

Results: This study comprised 516 patients, with a mean follow-up duration of 793.4±58.6 days. Following the integration of demographic data, laboratory results, and ABPM findings into the ML models, age, high-density lipoprotein cholesterol, triglycerides, and the standard deviation of systolic blood pressure (SDsis) were identified as predictors for future HT. A logistic regression with the selected variables (age, diabetes mellitus history, HDL, triglycerides, white blood cell count, and SDsis) using the full data set gave the following log odds 0.0737 (P<0.001), 0.7146 (P<0.001), -0.0160 (P=0.071), 0.0026 (P=0.002), 0.0857 (P=0.069), and 0.0850 (P=0.005), respectively. The corresponding probability values of age, diabetes mellitus history, HDL, triglycerides, white blood cell count, and SDsis were 0.5184, 0.6714, 0.4960, 0.5006, 0.5214, and 0.5212, respectively. This indicates a unit increase in all factors, except diabetes mellitus history, increases the probability of future HT by 50%. A history of diabetes, however, increases the probability of future HT by more than two thirds. The history of diabetes mellitus emerged as the most crucial predictor of future HT across all applied methods.

Conclusions: ML methods appear to be valuable tools for predicting future HT. The widespread adoption of these methods and the refinement of more comprehensive models will lay the groundwork for future studies.