The Association of Periodontal Inflammation and Systemic Health Indicators: A Machine Learning Approach

ABSTRACT Aim The relationship between oral and systemic inflammation has profound implications for understanding the broader health impacts of periodontitis. The aim of this study was to (a) explore the association between periodontal inflammation and markers of systemic inflammation and metabolic health, and (b) preliminarily assess periodontal status based on systemic health indicators using machine learning techniques. Methods Data from a cross‐sectional cohort ( N = 667) were modelled (simple/multiple linear, fractional polynomial, logistic and random forest regression) to examine the association between systemic and periodontal measures. Three classifiers—random forest (RF), support vector machine (SVM) and gradient boosting (GB)—were used using periodontal inflamed surface area (PISA) and demographic and anthropometric variables (age, gender, ethnicity, body mass index [BMI] and smoking habits) as inputs to predict systemic inflammation (defined using serum C‐reactive protein [CRP] levels). The best performing classification models (evaluated using area under the curve, AUC analyses) were validated using a second nationally representative dataset from the National Health and Nutrition Examination Surveys (NHANES) 2001–2002 and 2003–2004 combined datasets ( N = 2288). Next, RF, SVM and GB were employed incorporating a set of systemic parameters (including serum CRP and lipid profiles) to predict the diagnosis of periodontitis. The best performing classification models were then validated using the NHANES 2009–2010 ( N = 664) dataset. Results A nonlinear trend of CRP levels and PISA was confirmed by fractional polynomial regression ( p = 0.008). Further, multiple linear regression analyses (adjusted for age, gender, ethnicity, BMI and smoking habits) confirmed a statistically significant relationship between log‐transformed CRP levels and PISA ( p < 0.0001). Logistic regression confirmed a relationship between PISA and low‐density lipoprotein (LDL) in both crude and adjusted models. Among the classification models, SVM showed the highest performance in distinguishing CRP < 2 mg/L from CRP ≥ 2 mg/L (AUC = 0.71). The SVM model was successfully replicated in the NHANES 2001–2002 and 2003–2004 waves (AUC = 0.74). Prediction of periodontitis status (case vs. control) based on systemic indicators using the SVM model achieved the best performance with a mean AUC of 0.82. This was partially confirmed after external validation using the 2009–2010 NHANES dataset (AUC of 0.72). Conclusion This study confirmed a consistent relationship between measures of cumulative periodontal inflammation and systemic inflammation through machine learning models. A predictive model incorporating systemic health parameters helped in identifying a case with periodontitis. Both models have potential for use in primary healthcare settings, including screening programmes, as providing confirmation of the bidirectional link between periodontitis and systemic health.