Multifactorial contributors to knee osteoarthritis identified by Machine Learning and Bayesian network analysis: A cross-sectional study of the Iwaki Health Promotion Project

Abstract

Objective To explore factors associated with radiographic knee osteoarthritis (KOA) among a broad range of variables using community-based health examination data through machine-learning-based classification and Bayesian network (BN) analysis. Methods This cross-sectional study included 703 participants from the 2022 Iwaki Health Promotion Project. Radiographic KOA was defined as Kellgren–Lawrence grade ≥2 in the right knee. From 993 candidate variables encompassing clinical, imaging, laboratory, lifestyle, and dietary data, machine learning with feature selection was used to develop binary classification models for the overall population and, as a subgroup analysis, women. Model performance was evaluated using 10-fold cross-validation. BN analysis with 1000 bootstrap resamplings was subsequently performed to explore conditional dependence structures among the selected variables and KOA. Results Fifty variables in the overall population and 54 in women were retained after feature selection. The classification models achieved AUCs of 0.849 and 0.808, respectively, compared with 0.761 for a reference model including age, sex, and BMI. Variables related to knee symptoms, amino acids, fatty acids, sex hormones, and dietary habits were associated with KOA in the BN analyses. Dietary habit–related variables were particularly prominent in the overall population. The frequencies of carbonated beverage and deep-fried chicken consumption were significantly associated with KOA in logistic regression analysis. Conclusions Machine learning combined with BN analysis identified multidomain factors associated with radiographic KOA from comprehensive community health data. These findings are exploratory and require validation in longitudinal and independent cohorts.