AdaBoost-based differential diagnosis of gouty arthritis and rheumatoid arthritis: model construction, temporal validation, and SHAP visualization

Abstract

Background The phenotypic overlap between gouty arthritis (GA) and Rheumatoid Arthritis (RA) poses diagnostic challenges, particularly in resource-limited settings. This study aims to develop a robust, non-invasive machine learning (ML) framework based on routine hematological parameters to facilitate precise differential diagnosis and to evaluate its robustness in a temporally independent cohort. Methods Adopting a temporal validation design, we retrospectively analyzed clinical data from a tertiary hospital. A development cohort (2021–2025; n = 9,903) was utilized for model construction, while an independent historical cohort (2015–2020; n = 1,983) served as an time-based internal validation set. Feature selection was performed via LASSO regression, followed by benchmarking nine ML algorithms, including AdaBoost. Performance was evaluated using the Area Under the Curve (AUC) and Decision Curve Analysis (DCA), with interpretability enhanced by the SHAP framework. Results Six pivotal features were identified, comprising demographics, traditional markers, and novel composite ratios (CHOL/TG). AdaBoost demonstrated superior performance, achieving an AUC of 0.898 in the internal test set and an impressive 0.897 in the time-based internal validation, indicating exceptional generalization. SHAP analysis revealed that beyond conventional markers, CHOL/TG is a key predictor, revealing a unique interaction pattern of lipid metabolism that distinguishes gouty arthritis (GA) from rheumatoid arthritis (RA). Conclusion We successfully developed and validated an AdaBoost-based model leveraging routine blood indices for precise GA/RA differentiation. Validated across temporal cohorts, this cost-effective tool holds significant promise for optimizing early triage in primary care, while offering novel pathophysiological insights through the lens of lipid metabolism.