Clinical validation of machine learning models for periodontal disease using established metabolic risk factors

Aims: Artificial Intelligence (AI)-based models are increasingly applied in periodontal research; however, their clinical validity relative to classical statistical approaches remains insufficiently evaluated. This study was designed as a methodological validation exercise, using well-established metabolic risk factors for periodontitis to benchmark model behavior against known biological relationships. It aimed to compare classical statistical models and AI-based analytical approaches in their ability to reproduce established associations between metabolic markers, hemoglobin A1c (HbA1c), and body-mass index (BMI), and periodontal outcomes.Methods: A cross-sectional methodological comparison was conducted using data from 1,852 adults in the NHANES 2009-2010 dataset. Periodontal pocket depth (PD) and periodontal disease status were modeled using classical linear and logistic regression, AI-optimized linear regression, random forest, deep learning, and a pseudo-multi-state logistic model (MSM). Models were evaluated with respect to stability, interpretability, and biological plausibility, in addition to numerical performance.Results: HbA1c demonstrated a modest but statistically significant association with PD and disease probability across all modeling approaches, whereas BMI showed a statistically detectable but clinically negligible contribution. Classical linear regression and AI-assisted linear models yielded stable, clinically interpretable results. Deep learning achieved slightly improved numeric performance but did not reveal additional nonlinear structure. Random forest models showed unstable and clinically implausible prediction patterns. Probability-based models consistently demonstrated a monotonic increase in periodontal disease risk with increasing HbA1c.Conclusion: For biologically modest and predominantly linear metabolic-periodontal relationships, traditional statistical models remain robust, interpretable, and clinically reliable. AI-based methods offer limited added value beyond incremental numerical refinement, underscoring the importance of aligning analytical complexity with biological signal strength when validating analytical approaches in periodontal epidemiology.