DAP5: a composite metabolic index for predicting diabetes–arthritis comorbidity—development, external validation, and comparison with machine learning

Type 2 diabetes and arthritis frequently coexist and are thought to interact bidirectionally through metabolic inflammation, yet few simple indices are available to predict their co-occurrence. This study developed a routine blood-based composite marker DAP5 (Diabetes-Arthritis Prediction Factor 5) and evaluated its predictive ability in two distinct ethnic cohorts. A total of 5,733 disease-free middle-aged and older participants from the China Health and Retirement Longitudinal Study (CHARLS) 2011–2018 were included as the model development cohort. Five core indicators were pre-specified on pathophysiological grounds, and their independent associations were verified by univariate regression and receiver operating characteristic (ROC) analysis to construct the DAP5 index. Multivariate logistic regression, restricted cubic spline, and bootstrap resampling were used to comprehensively evaluate its predictive performance and robustness. An Extreme Gradient Boosting (XGBoost) machine learning model was also constructed for head-to-head comparison, and cross-ethnic external validation was performed in the English Longitudinal Study of Ageing (ELSA). The DAP5 index consists of five routine physical examination indicators: glycated hemoglobin (HbA1c), fasting blood glucose (FBG), triglycerides (TG), body mass index (BMI), and high-density lipoprotein cholesterol (HDL-C). The unadjusted AUROC of DAP5 was 0.723 (95% CI: 0.676–0.770) in the full CHARLS development cohort ( n = 5,733), comparable to the strongest single indicator HbA1c (0.723 vs 0.725), numerically higher than FBG, BMI, TG and HDL-C, and comparable to the classic insulin resistance index Metabolic Score for Insulin Resistance (METS-IR) ( p = 0.219); that is, DAP5 was not superior to HbA1c alone in terms of predictive discrimination, and its practical value lies in integrating five routinely available indicators into a single risk-ranking measure rather than in improved discrimination. After multivariable adjustment, the AUROC was 0.744. After adjusting for confounding factors including age, sex, and lifestyle, each 1-unit increase in DAP5 was associated with significantly higher odds of comorbidity (OR = 5.933, 95% CI: 3.978–8.831, p < 0.001). The DAP5 index and the XGBoost model achieved AUROCs of 0.755 and 0.670, respectively, in the independent ELSA external validation cohort ( n = 4,489), providing preliminary evidence for the generalizability of the core metabolic indicators, whereas the prespecified cut-off was not supported in the external cohort (sensitivity = 0%) and requires further evaluation before broader application. In participants free of arthritis at baseline, DAP5 was not associated with incident arthritis (OR = 1.04, 95% CI: 0.87–1.24, p = 0.643; AUROC = 0.582). An exploratory mediation analysis, unstable given the low event rate, was consistent with an association operating mainly through incident diabetes, although the indirect effect itself was not statistically significant (total effect: OR = 5.86, 95% CI: 3.89–8.82). The exploratory mediation and four-category analyses were consistent with the index primarily capturing the shared dysglycaemic substrate underlying both conditions rather than serving as an arthritis-specific predictor.