Objective:

To improve early prediction of rheumatoid arthritis (RA) by utilizing routine laboratory parameters and machine learning techniques, specifically targeting anti-CCP positivity and early RA onset.

Approach:

• Study Design: 500 patients were enrolled, and 29 routine laboratory features were collected to predict anti-CCP positivity and early RA onset, defined as dual binary prediction targets.

Key Findings:

• Logistic regression achieved the best overall performance with accuracy of 0.848, AUC of 0.857, F1 of 0.910, and MCC of 0.441.
• The Transformer deep learning model performed well with an AUC of 0.812.
• ESR and CRP were identified as the most important positive predictive drivers, while albumin was a key protective factor.

Interpretation:

The machine learning pipeline predicts early RA risk, and the SHAP analysis provides interpretable decision rationale.

Limitations:

• Existing studies often evaluate only a single or few models, lacking systematic cross-comparison.
• The 'black-box' nature of deep learning models limits clinical credibility.
• RA datasets commonly exhibit class imbalance, which can affect model performance.

Conclusion:

The study demonstrates the potential of machine learning and SHAP analysis in enhancing early RA risk prediction using routine laboratory data.