Objectives Connective tissue disease–associated interstitial lung disease (CTD-ILD) exhibits heterogeneous clinical outcomes, and traditional ILD-GAP score provides limited prognostic precision. We hypothesized that integration of imaging-derived fibrosis measures would improve prognostic discrimination compared with physiology-based models. We aimed to develop and externally validate integrated prognostic systems for risk stratification in CTD-ILD. Method In this multicenter retrospective study, patients with CTD-ILD confirmed by multidisciplinary evaluation were enrolled from two tertiary centers. Baseline demographic, functional, laboratory, and imaging data were collected at diagnosis. Three prognostic systems were evaluated: System A (ILD-GAP and fibrosis score independently), system B (a composite model integrating GAP and fibrosis scores), and system C (machine learning models using clinical and imaging variables). The primary endpoint was a composite adverse outcome during follow-up. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC) with internal and external validation. Results In external validation, the fibrosis score demonstrated stronger discrimination than the ILD-GAP index (AUC 0.798 vs. 0.768). The composite model showed modest improvement (AUC 0.821). Machine learning models achieved the highest discrimination, with random forest and support vector machine yielding AUCs of 0.833 (95% CI 0.741–0.927) and 0.826 (95% CI 0.735–0.910), respectively. SHAP analysis identified fibrosis extent, DLCO, and age as key contributors to prediction. Conclusions In CTD-ILD, imaging-informed and data-driven models improve prognostic discrimination compared with traditional physiology-based indices. The fibrosis score plays a central role in outcome prediction, while composite and machine learning approaches provide incremental refinement for individualized risk stratification. <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 • Imaging-derived fibrosis score demonstrated superior prognostic discrimination compared with the traditional ILD-GAP index in CTD-ILD, highlighting the central role of structural lung damage in outcome prediction. • Integration of clinical, functional, and HRCT-derived variables through optimized composite modeling significantly improved risk stratification, with machine learning approaches (random forest and support vector machine) achieving the highest predictive performance. • SHAP analysis identified fibrosis extent, DLCO, and age as the dominant contributors to adverse outcomes, underscoring the complementary value of imaging and physiology in individualized risk assessment. </entry> </row> </tbody> </tgroup> </Table>

Integrated imaging-informed and machine learning–based prognostic models for risk stratification in connective tissue disease–associated interstitial lung disease: a multicenter study

期刊论文
发表日期：

2026年5月11日

查看全文

Yao Xu,
Jinpeng Hou,
Dong Yan,
Wen Zhu,
Dan Jin,
Wei Zhang,
Tian Ren,
Guohua Fan

摘要

Objectives

Connective tissue disease–associated interstitial lung disease (CTD-ILD) exhibits heterogeneous clinical outcomes, and traditional ILD-GAP score provides limited prognostic precision. We hypothesized that integration of imaging-derived fibrosis measures would improve prognostic discrimination compared with physiology-based models. We aimed to develop and externally validate integrated prognostic systems for risk stratification in CTD-ILD.

Method

In this multicenter retrospective study, patients with CTD-ILD confirmed by multidisciplinary evaluation were enrolled from two tertiary centers. Baseline demographic, functional, laboratory, and imaging data were collected at diagnosis. Three prognostic systems were evaluated: System A (ILD-GAP and fibrosis score independently), system B (a composite model integrating GAP and fibrosis scores), and system C (machine learning models using clinical and imaging variables). The primary endpoint was a composite adverse outcome during follow-up. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC) with internal and external validation.

Results

In external validation, the fibrosis score demonstrated stronger discrimination than the ILD-GAP index (AUC 0.798 vs. 0.768). The composite model showed modest improvement (AUC 0.821). Machine learning models achieved the highest discrimination, with random forest and support vector machine yielding AUCs of 0.833 (95% CI 0.741–0.927) and 0.826 (95% CI 0.735–0.910), respectively. SHAP analysis identified fibrosis extent, DLCO, and age as key contributors to prediction.

Conclusions

In CTD-ILD, imaging-informed and data-driven models improve prognostic discrimination compared with traditional physiology-based indices. The fibrosis score plays a central role in outcome prediction, while composite and machine learning approaches provide incremental refinement for individualized risk stratification.

Key Points

• Imaging-derived fibrosis score demonstrated superior prognostic discrimination compared with the traditional ILD-GAP index in CTD-ILD, highlighting the central role of structural lung damage in outcome prediction.

• Integration of clinical, functional, and HRCT-derived variables through optimized composite modeling significantly improved risk stratification, with machine learning approaches (random forest and support vector machine) achieving the highest predictive performance.

• SHAP analysis identified fibrosis extent, DLCO, and age as the dominant contributors to adverse outcomes, underscoring the complementary value of imaging and physiology in individualized risk assessment.