A75-14 Inflammation-Derived Indices Enhance the Identification of Acute Exacerbation in Idiopathic Pulmonary Fibrosis: A Dual-Center Machine Learning Study

Background Acute exacerbation (AE) is a devastating complication of idiopathic pulmonary fibrosis (IPF), yet reliable tools for early identification and risk stratification remain limited. While systemic inflammation is implicated in AE-IPF pathogenesis, the predictive value of inflammation-derived indices remains underexplored. We aimed to investigate the association between inflammation-derived indices and AE-IPF and to develop an interpretable machine learning (ML) model for AE identification. Methods In this dual-center cross-sectional study, consecutive IPF patients were enrolled from two tertiary hospitals between October 2021 and October 2025. AE-IPF was strictly defined according to the 2018 ATS/ERS/JRS/ALAT guideline, requiring acute radiological deterioration and exclusion of alternative causes, including infection and heart failure. Data from two centers were pooled to maximize statistical power. Nineteen inflammation-derived indices were calculated. Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to select key inflammatory features. Associations were examined using univariable and multivariable logistic regression, subgroup analyses, sensitivity analyses, and restricted cubic spline modeling. Multiple ML algorithms were developed with 10-fold cross-validation. Model performance was assessed using discrimination, calibration, and overall accuracy, and model interpretability was evaluated using Shapley Additive Explanations (SHAP). Results A total of 489 patients were included, of whom 157 experienced AE. LASSO regression consistently identified the neutrophil-to-high-density lipoprotein ratio (NHR), platelet-to-high-density lipoprotein ratio (PHR), and atherogenic index of plasma (AIP) as key inflammation-derived predictors. These indices were independently and robustly associated with AE-IPF across multivariable analyses, subgroup and sensitivity analyses, and nonlinear modeling (e.g., AIP: OR 4.44, 95% CI 2.25-8.78; all P < 0.001). Among the ML models, the XGBoost model demonstrated superior performance. Specifically, the integrated model incorporating both inflammatory and clinical features achieved an AUC of 0.805 in the testing cohort, significantly outperforming the inflammation-only model (AUC 0.758) and the baseline clinical model (AUC 0.697). The final model showed high specificity (0.889), good overall accuracy (0.816), and favorable calibration (Brier score 0.249). SHAP analysis visualized the decision-making process, revealing that elevated AIP, PHR, and NHR, alongside key clinical features (lower FVC, honeycombing), were the most influential contributors to identifying AE-IPF. Conclusions Inflammation-derived indices are independently and robustly associated with AE-IPF and substantially enhance risk stratification and identification when integrated into machine learning models. Beyond providing a practical tool for precision management, our findings highlight the potential interplay between lipid metabolism and neutrophilic inflammation in the pathogenesis of AE-IPF, offering novel insights for future therapeutic targeting.