Objective This study constructed a predictive model for interstitial lung disease (ILD) in patients with rheumatoid arthritis (RA) and explored the value of interpretable machine learning. Methods The medical records of 400 hospitalized RA patients treated in the Department of Rheumatology and Immunology, the First Affiliated Hospital of Anhui University of Chinese Medicine, from July 2021 to January 2025 were collected retrospectively. The patients were randomly allocated into a model training set (n = 280) and a validation set (n = 120) at a ratio of 7:3. Important feature variables were screened using the least absolute shrinkage and selection operator (LASSO) regression. The predictive models for ILD in RA patients were constructed using four machine learning algorithms: light gradient boosting machine (LightGBM), logistic regression (LR), extreme gradient boosting (XGBoost), and support vector machine (SVM). The performance of these models was measured based on the area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA). The Shapley additive explanation (SHAP) algorithm was applied to conduct interpretability analysis for the optimal model. Results In this study, the prevalence of ILD was found to be 15% (60/400). LASSO regression was used to select six feature variables: age, albumin, rheumatoid factor, anti-cyclic citrullinated peptide (CCP) antibody, high-sensitivity C-reactive protein, and leflunomide history. The XGBoost model displayed improved performance on the validation set, achieving an AUC of 0.841, along with a 0.937 accuracy and 0.899 F1-score, respectively. DCA also demonstrated favorable performance. Using the SHAP algorithm, the order of importance of the feature variables was age, albumin, rheumatoid factor, anti-CCP antibody, high-sensitivity C-reactive protein, and leflunomide history. Conclusions In conclusion, age, albumin, rheumatoid factor, anti-CCP antibody, high-sensitivity C-reactive protein, and leflunomide medication history are the key clinical predictors for RA-ILD. The XGBoost model combined with SHAP interpretation shows good predictive performance and clinical interpretability, which can assist clinicians in early risk assessment of RA-ILD.<Table Float="No" ID="Taba"> <tgroup cols="2"> <colspec align="left" colname="c1" colnum="1" /> <colspec align="left" colname="c2" colnum="2" /> <tbody> <row> <entry align="left" nameend="c2" namest="c1"> Key Points • Age, albumin, rheumatoid factor, anti-CCP antibody, high-sensitivity C-reactive protein, and leflunomide medication history are the key clinical predictors for RA-ILD. • The XGBoost model achieved superior predictive performance (AUC = 0.841) and outperformed LightGBM, LR, and SVM for RA-ILD risk stratification. • The SHAP algorithm was innovatively applied to achieve full model interpretability, clearly quantifying the contribution and effect direction of each risk factor. </entry> </row> </tbody> </tgroup> </Table>

Development of a risk prediction model for rheumatoid arthritis-associated interstitial lung disease based on interpretable machine learning

期刊论文
发表日期：

2026年7月22日

查看全文

Xuanbin Li,
Juan Yuan,
Chunbiao Li,
Man Luo,
Yunxiang Cao,
Jianting Wen,
Yumei Wu

摘要

Objective

This study constructed a predictive model for interstitial lung disease (ILD) in patients with rheumatoid arthritis (RA) and explored the value of interpretable machine learning.

Methods

The medical records of 400 hospitalized RA patients treated in the Department of Rheumatology and Immunology, the First Affiliated Hospital of Anhui University of Chinese Medicine, from July 2021 to January 2025 were collected retrospectively. The patients were randomly allocated into a model training set (n = 280) and a validation set (n = 120) at a ratio of 7:3. Important feature variables were screened using the least absolute shrinkage and selection operator (LASSO) regression. The predictive models for ILD in RA patients were constructed using four machine learning algorithms: light gradient boosting machine (LightGBM), logistic regression (LR), extreme gradient boosting (XGBoost), and support vector machine (SVM). The performance of these models was measured based on the area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA). The Shapley additive explanation (SHAP) algorithm was applied to conduct interpretability analysis for the optimal model.

Results

In this study, the prevalence of ILD was found to be 15% (60/400). LASSO regression was used to select six feature variables: age, albumin, rheumatoid factor, anti-cyclic citrullinated peptide (CCP) antibody, high-sensitivity C-reactive protein, and leflunomide history. The XGBoost model displayed improved performance on the validation set, achieving an AUC of 0.841, along with a 0.937 accuracy and 0.899 F1-score, respectively. DCA also demonstrated favorable performance. Using the SHAP algorithm, the order of importance of the feature variables was age, albumin, rheumatoid factor, anti-CCP antibody, high-sensitivity C-reactive protein, and leflunomide history.

Conclusions

In conclusion, age, albumin, rheumatoid factor, anti-CCP antibody, high-sensitivity C-reactive protein, and leflunomide medication history are the key clinical predictors for RA-ILD. The XGBoost model combined with SHAP interpretation shows good predictive performance and clinical interpretability, which can assist clinicians in early risk assessment of RA-ILD.

Key Points

• Age, albumin, rheumatoid factor, anti-CCP antibody, high-sensitivity C-reactive protein, and leflunomide medication history are the key clinical predictors for RA-ILD.

• The XGBoost model achieved superior predictive performance (AUC = 0.841) and outperformed LightGBM, LR, and SVM for RA-ILD risk stratification.

• The SHAP algorithm was innovatively applied to achieve full model interpretability, clearly quantifying the contribution and effect direction of each risk factor.