# A Cross-Sectional Study Based on Deep Learning to Explore the Effect of Triglyceride/Glucose Index on Periodontitis: An Analysis Based on the Large NHANES Database

## TL;DR

This study uses deep learning on a large health database to show that the TyG index, a marker of insulin resistance, is linked to periodontitis risk.

## Contribution

The study introduces a deep learning model to demonstrate the predictive value of the TyG index for periodontitis.

## Key findings

- The TyG index was the most influential predictor in the deep learning model for periodontitis.
- The model achieved an AUC-ROC of 0.7482 and 87.13% accuracy in predicting periodontitis.
- A dose–response relationship was found between TyG index and periodontitis risk.

## Abstract

Periodontitis is a common chronic inflammatory disease closely associated with metabolic syndrome. The triglyceride-glucose (TyG) index is a surrogate marker of insulin resistance. However, its relationship with periodontitis remains underexplored. This study aims to utilise the large national database (NHANES) and explore the predictive value of TyG index for periodontitis through a deep learning model, and to clarify its correlation.

This study utilised data from the National Health and Nutrition Examination Survey (NHANES), including participants with complete demographic, laboratory, and oral health data. TyG index was calculated and incorporated into a deep learning model. A neural network with multiple hidden layers was built using PyTorch framework and trained using binary cross-entropy loss and the Adam optimiser. Model performance in predicting periodontitis was evaluated by metrics including -ROC, accuracy, sensitivity, and specificity. Traditional logistic regression and SHapley Additive ex-Planations (SHAP) algorithms were applied to validate and interpret the model’s predictive capacity and feature contributions.

2,834 participants were included. Baseline characteristics showed that individuals with periodontitis had significantly higher age, BMI, fasting glucose, and triglyceride levels compared to those without periodontitis (P < 0.001). The deep learning model demonstrated good performance on the test set, with AUC-ROC of 0.7482 and an accuracy of 87.13%. Feature importance analysis revealed that TyG index was the most influential predictor in the model. Logistic regression analysis indicated a significant dose–response relationship between TyG index and risk of periodontitis. Although statistical significance decreased after full adjustment for confounders, trend analysis still supported the TyG index as a potential independent risk factor.

The TyG index was significantly dose-dependent and correlated with the risk of periodontitis, and the deep learning model demonstrated excellent predictive performance. The TyG index may serve as a simple and low-cost biomarker for the risk stratification and early identification of periodontitis, providing new ideas for the interdisciplinary management of oral-metabolic diseases.