Periodontitis Risk Prediction

Interpretable machine learning that predicts gum-disease severity from information collected outside the dental chair — trained on 10,683 fully examined NHANES participants and explained, factor by factor, with SHAP.

Open Live App

Live at nhanes-oral-health-ml.vercel.app

Core Capabilities

An Outcome You Can Trust

The target label is not a proxy. Every participant is graded with the CDC/AAP case definition from real full-mouth probing depth and clinical attachment loss measurements. Survey-weighted prevalence reproduces the published national estimates exactly — 42.2% any periodontitis, 7.8% severe — which is how we know the pipeline is right before a single model is trained.

NHANES 2009–2014, the three cycles with a full-mouth exam
30,468 participants screened down to a 10,683 final cohort
Prevalence matches Eke et al. to the decimal point

Simple Models, Measured Honestly

Multinomial logistic, ordinal proportional-odds and XGBoost are compared against an age-only baseline on a held-out test set of 2,137 participants, with 95% bootstrap confidence intervals on every number. Nothing from the dental examination is ever used as a predictor, so no leakage inflates the result.

Macro AUC 0.746 against a 0.628 age-only baseline
AUC around 0.80 for moderate and severe disease
Logistic regression matches XGBoost — transparency for free

Explanations That Agree With the Statistics

TreeSHAP gives global importance, per-class effects and an individual explanation for any person the model scores. Those rankings are then checked against classical odds ratios from the ordinal model — and they agree, which is what turns an explanation into evidence rather than decoration.

Strongest risk: age (OR 2.02), current smoking (1.29), HbA1c (1.13)
Protective: female sex (0.68), education (0.77), income (0.79)
SHAP and odds ratios rank the same drivers at the top

Why? (SHAP)

Age 61 → pushes severe risk up

Current smoker → pushes up

College degree → pushes down

What You Can Do in the Live App

Risk Predictor

Enter a profile, or load a real held-out NHANES participant and compare the prediction against their actual examination result. All seven models can be run side by side.

Case Definition Calculator

Edit probing depth and attachment loss tooth by tooth and watch the CDC/AAP class change, along with the exact criteria that were met.

Model Performance

Every metric with its confidence interval, odds-ratio forest plots, confusion matrices, calibration curves and the full SHAP figure set.

Fully Reproducible

One command downloads the CDC data, rebuilds the cohort, retrains and re-evaluates. Unit tests assert the deployed models match the originals.

Population Health Research, Not a Diagnostic Tool

This is a cross-sectional study of US adults aged 30 and over. It surfaces population-level associations and relative risk rankings — it does not diagnose individuals. If you need interpretable ML built on your own clinical or survey data, we should talk.

Discuss a Healthcare AI Project