Background Rheumatoid arthritis (RA) is a chronic autoimmune disease that can cause significant disability. Early detection is vital for improving outcomes, but current diagnostic methods are costly and lack accessibility for wide-scale screening. While routine blood cell parameters are promising, novel cellular morphological indices (e.g., NE-SFL, LY-Y) remain under-explored for RA screening. This study aimed to develop a machine learning model for RA screening by integrating novel morphological with routine blood parameters. Methods This retrospective study analyzed 43 blood cell parameters from 3009 participants. Feature selection used multiple machine learning (ML) algorithms, and eight predictive models were developed. Performance was evaluated on an independent test set via AUC, accuracy, sensitivity, specificity, calibration, and decision curve analysis. Model interpretability was provided by SHapley Additive exPlanations (SHAP). Results Seven key predictors were identified. The XGBoost model performed best, achieving an AUC of 0.929, accuracy of 0.890, sensitivity of 0.774, and specificity of 0.915 on the test set. SHAP analysis showed that novel cellular morphological parameters were the primary predictive drivers, exceeding some traditional inflammatory markers. Conclusion We developed and validated a high-performance, interpretable ML model using blood parameters to identify individuals at high risk for RA. The results highlight the significant value of novel morphological parameters for RA screening. An accompanying online tool offers a feasible, low-cost solution for non-invasive RA screening in primary care settings.<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 •Focus on RA screening: The model is designed specifically for screening and risk stratification of RA where timely intervention is most critical. •High performance with routine data: Using only seven readily available blood parameters, our XGBoost model achieved an AUC of 0.929 on an independent test set, demonstrating performance comparable to more complex or costly modalities •Translational readiness: We have operationalized this model into a freely accessible online tool, enabling immediate clinical use. This tool can serve as an efficient “pre-screening filter” in routine health checks, guiding the referral of high-risk individuals for specialist evaluation and confirmatory testing (e.g., anti-CCP, imaging). •Interpretability for clinical trust: We employed SHAP analysis to ensure model transparency, explaining both global feature importance and individual predictions. This interpretability is crucial for fostering clinician adoption and understanding the biological underpinnings of the predictions, such as the prominent role of cellular morphological parameters like LY-Y and NE-SFL. </entry> </row> </tbody> </tgroup> </Table>

An interpretable machine learning tool for rheumatoid arthritis screening: integrating novel cellular morphological parameters with routine blood count indices

期刊论文
发表日期：

2026年4月13日

查看全文

Lan You,
Xin Li,
Ruocheng Luo,
Chen Peng,
Qianhui Liu,
Ruichun Sun,
Nan Zhang,
Jie Hou,
Bin Yang

摘要

Background

Rheumatoid arthritis (RA) is a chronic autoimmune disease that can cause significant disability. Early detection is vital for improving outcomes, but current diagnostic methods are costly and lack accessibility for wide-scale screening. While routine blood cell parameters are promising, novel cellular morphological indices (e.g., NE-SFL, LY-Y) remain under-explored for RA screening. This study aimed to develop a machine learning model for RA screening by integrating novel morphological with routine blood parameters.

Methods

This retrospective study analyzed 43 blood cell parameters from 3009 participants. Feature selection used multiple machine learning (ML) algorithms, and eight predictive models were developed. Performance was evaluated on an independent test set via AUC, accuracy, sensitivity, specificity, calibration, and decision curve analysis. Model interpretability was provided by SHapley Additive exPlanations (SHAP).

Results

Seven key predictors were identified. The XGBoost model performed best, achieving an AUC of 0.929, accuracy of 0.890, sensitivity of 0.774, and specificity of 0.915 on the test set. SHAP analysis showed that novel cellular morphological parameters were the primary predictive drivers, exceeding some traditional inflammatory markers.

Conclusion

We developed and validated a high-performance, interpretable ML model using blood parameters to identify individuals at high risk for RA. The results highlight the significant value of novel morphological parameters for RA screening. An accompanying online tool offers a feasible, low-cost solution for non-invasive RA screening in primary care settings.

Key Points

•Focus on RA screening: The model is designed specifically for screening and risk stratification of RA where timely intervention is most critical.

•High performance with routine data: Using only seven readily available blood parameters, our XGBoost model achieved an AUC of 0.929 on an independent test set, demonstrating performance comparable to more complex or costly modalities

•Translational readiness: We have operationalized this model into a freely accessible online tool, enabling immediate clinical use. This tool can serve as an efficient “pre-screening filter” in routine health checks, guiding the referral of high-risk individuals for specialist evaluation and confirmatory testing (e.g., anti-CCP, imaging).

•Interpretability for clinical trust: We employed SHAP analysis to ensure model transparency, explaining both global feature importance and individual predictions. This interpretability is crucial for fostering clinician adoption and understanding the biological underpinnings of the predictions, such as the prominent role of cellular morphological parameters like LY-Y and NE-SFL.