# Multiple machine-learning-driven metabolic frameworks for long-term prognostic risk assessment in patients with coexisting hypertension and obstructive sleep apnea:insights from a multicenter cohort study

## TL;DR

This study uses machine learning to better predict cardiovascular risks in patients with hypertension and sleep apnea, using more detailed metabolic indicators than BMI alone.

## Contribution

The novel use of multiple machine learning models and new metabolic indices improves long-term risk prediction for patients with hypertension and obstructive sleep apnea.

## Key findings

- The XGBoost model achieved an AUC of 0.898 for predicting major adverse cardiovascular and cerebrovascular events.
- Metabolic indices like TyG-BMI and TyG-WHtR were identified as key predictors alongside traditional clinical variables.
- Machine learning models outperformed conventional methods in predicting risk for high-risk patients.

## Abstract

Predictive obesity indices are often based on the body mass index (BMI). Although BMI is widely used, it does not provide a direct measure of obesity. We aimed to utilize multiple machine learning-driven metabolic frameworks to investigate the long-term risk of major adverse cardiovascular and cerebrovascular events (MACCEs) in individuals with hypertension and obstructive sleep apnea (OSA).

This study included 708 patients with hypertension and OSA between January 2017 and December 2021. The measurements of height, weight, neck circumference (NC), waist circumference (WC), neck-circumference-to-height ratio (NHtR), and waist-to-height ratio (WHtR) were collected to calculate the triglyceride-glucose (TyG)-BMI, as well as TyG-NC, TyG-WC, TyG-NHtR, and TyG-WHtR indices.

All patients were allocated to the training cohort (n = 446) and independent validation cohort (n = 262). The Boruta plot presented for identifying key predictors is as follows: male sex, age, TyG, TyG-BMI, HbA1c, FPG, triglyceride, creatinine, fibrinogen and AHI. We constructed nine machine learning models-XGBoost, Light Gradient Boosting Machine, Random Forest, Decision Tree, Gradient Boosting, Multi-Layer Perceptron, Support Vector Machine, K-Nearest Neighbors, and Gaussian Naive Bayes-to predict MACCEs. The XGBoost model was selected due to its superior performance evidenced by an AUC of 0.898 (95% CI: 0.822–0.973) and net clinical benefit. SHAP analysis further clarified variable contributions to MACCE risk.

This study employed various machine-learning techniques and multidimensional data assessment, allowing for enhanced prediction of metabolic results and supporting the timely detection of high-risk patients with OSA and hypertension in need of focused preventive measures.

https://www.chictr.org.cn/bin/project/edit?pid=206415, identifier ChiCTR2300075727.