A75-19 A Novel Combined Fairfax IPF Clinical Score and Deep Learning UIP Classifier Model for Diagnosing Idiopathic Pulmonary Fibrosis

Rationale

Idiopathic Pulmonary Fibrosis (IPF) is a progressive fibrotic lung disease defined by the usual interstitial pneumonia (UIP) pattern and carries high mortality. Accurate diagnosis requires integrating clinical and radiologic data, and multidisciplinary discussion (MDD) is considered the diagnostic gold standard. However, many centers lack access to subspecialized pulmonologists and chest radiologists needed to conduct MDDs. To address this gap, we evaluated a combined diagnostic model incorporating the Fairfax IPF Clinical Score (FICS) and a deep learning-based radiomic classifier that identifies UIP pattern on CT (IQ-UIP). FICS demonstrates high sensitivity but low specificity for IPF, while IQ-UIP shows high sensitivity and specificity for detecting UIP pattern. We hypothesized that integrating these complementary clinical and radiologic predictors would improve discrimination between IPF and non-IPF ILD compared with the FICS score alone.

Methods

We conducted a retrospective, observational study at the University of California, San Diego ILD clinic. Seventy-nine patients with interstitial lung disease underwent FICS scoring and high- resolution CT analysis using the IQ-UIP algorithm. An expert MDD including ILD pulmonologist and chest radiologist established the reference diagnosis for each patient. Diagnostic performance for identifying IPF versus non- IPF was assessed for FICS alone, IQ-UIP alone, and a combined continuous model defined as the sum of standardized FICS and IQ-UIP probability scores. Receiver operating characteristic (ROC) curves were generated, and area under the curve (AUC) values were estimated with 95% confidence intervals. Optimal thresholds for binary classification were determined using the Youden J statistic. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated at cohort prevalence and at clinically relevant external prevalence estimates.

Results

Out of 79 patients, 20 were diagnosed with IPF by MDD. The combined continuous model demonstrated improved diagnostic performance compared with either component alone, with sensitivity 0.80, specificity 0.64, PPV 0.44, NPV 0.89, and AUC 0.79 for distinguishing IPF from non-IPF relative to MDD. Both FICS and IQ-UIP individually showed lower overall accuracy and AUC. We plan to further validate these findings in a larger IPF enriched cohort.

Conclusions

Integrating clinical risk stratification (FICS) with a radiomic UIP classifier (IQ-UIP) improves diagnostic discrimination for IPF compared with the FICS score alone. This combined clinical-radiographic model may offer a practical and scalable alternative to MDD in settings without access to subspecialty expertise, potentially expanding timely and accurate IPF diagnosis.

This abstract is funded by: Academy of Clinician Scholars Awards Committee University of California San Diego