Early prediction of bone destruction in rheumatoid arthritis through machine learning analysis of plasma metabolites

Abstract

Background: To develop a predictive model for bone destruction in patients with rheumatoid arthritis (RA), based on the characteristics of plasma metabolites and common clinical indicators.

Methods: The cohort comprised 60 patients with RA, with baseline metabolite features identified using the liquid chromatograph-mass spectrometer system. Radiographic outcomes were assessed using the van der Heijde-modified total Sharp score (mTSS) following a one-year follow-up period to quantify bone destruction. The longitudinal association between metabolites and radiographic progression was analyzed using several machine learning algorithms, and the significance of core metabolites was calculated. A new model incorporating metabolites and clinical indicators was created to evaluate its predictive performance for radiographic progression; the model was compared with other prediction models.

Results: The median increase in mTSS was 3.50. Of the 774 detected metabolites, 77 differed between patients with different outcomes. Core metabolites identified using the Gaussian Naive Bayes algorithm included mangiferic acid, O-acetyl-L-carnitine, 5,8,11-eicosatrienoic acid, and 16-methylheptadecanoic acid. A standardized bone erosion risk score (BERS) was developed based on these core metabolite features for assessing the radiographic progression outcome. Individuals with a high BERS exhibited a lower risk of rapid radiographic progression than those with a lower score (OR = 0.01, 95% CI = 0.01-0.03, P = 0.003). The "China-Japan Friendship Hospital-BERS Model" (CjBM), combining BERS with clinical features (methotrexate and C-reactive protein), produced an area under the receiver operating characteristic curve of 0.800. Moreover, compared with the reported models, the CjBM showed near statistical significance in identifying rapid radiographic progression; adding BERS can improve the discrimination of the original reported model (P_DeLong=0.035).

Conclusions: The CjBM was developed for early prediction of bone destruction in patients with RA, and the evaluation of BERS emphasizes the significance of metabolite features.

Keywords: Bone destruction; Machine learning; Metabolite; Prediction model; Rheumatoid arthritis.