# Risk Prediction of Edentulism in Chinese Adults: Insights From the China Health and Retirement Longitudinal Study (CHARLS)

## TL;DR

This study develops a risk prediction model for tooth loss in Chinese adults using health data, finding that age and inflammation are key factors.

## Contribution

The study introduces a validated logistic regression model for predicting edentulism using routinely available clinical data in Chinese adults.

## Key findings

- Logistic regression showed the best discrimination with an AUC of 0.695 for predicting edentulism.
- Age and C-reactive protein were the strongest predictors of tooth loss.
- The model demonstrated good calibration and potential for clinical use in risk stratification.

## Abstract

Edentulism concentrates in older adults and is associated with poorer diet, elevated malnutrition risk, and adverse health trajectories. Interpretable, well-calibrated tools applicable in routine care are needed to identify those at risk and support timely, tooth-preserving management.

We analyzed data from the nationally representative China Health and Retirement Longitudinal Study with baseline in 2011 to 2012 and follow-ups in 2013, 2015, and 2018. Adults with baseline dentition status and all prespecified predictors were eligible. Incident edentulism was ascertained at each wave. Predictors defined a priori included age, sex, body mass index, blood pressure, lipid profile, log-transformed C-reactive protein (ln-CRP), and estimated glucose disposal rate (eGDR). Four algorithms were trained in a development set and evaluated in an independent validation set: multivariable logistic regression, LASSO-penalized logistic regression, random forest, and extreme gradient boosting. Validation included area under the receiver-operating-characteristic curve (AUC) with nonparametric comparisons, calibration with logistic recalibration, decision-curve analysis across clinically relevant thresholds, and time-dependent AUC over follow-up.

Of 6130 adults, 5231 were dentate at baseline and formed the risk set; incident edentulism occurred in 2013 (n = 338), 2015 (n = 253), and 2018 (n = 308). Logistic regression yielded the highest discrimination (AUC 0.695) with stable performance across waves and over 2-7 years. Calibration closely matched the ideal after simple recalibration, and decision curves showed consistent net benefit versus treat-all and treat-none strategies. In the final model, age and C-reactive protein (ln-CRP) were dominant independent predictors; men had lower risk, and total cholesterol showed a modest inverse association. A nomogram was derived to enable point-of-care risk stratification.

In a large, nationally representative cohort, an internally validated logistic regression model based on routinely available data showed moderate discrimination, good calibration, and potential clinical utility for predicting incident edentulism. Age and low-grade systemic inflammation were the main contributors to risk stratification and may help guide early, tooth-preserving care.