A20-07 Identifying Preoperative CT Biomarkers Predictive of Chronic Lung Allograft Dysfunction Through a Systems Radiology Approach

Abstract Rationale Chronic lung allograft dysfunction (CLAD) remains the leading cause of late graft failure and mortality after lung transplantation, affecting over 50% of recipients within five years. Current diagnosis relies on spirometry, which detects decline only after irreversible injury. Early, noninvasive risk stratification remains an unmet clinical need. Using a systems radiology approach that integrates quantitative CT-derived lung morphology, cardiopulmonary vasculatures, and body composition features with clinical data, we hypothesized that preoperative imaging phenotypes could identify recipients at risk for future CLAD before functional decline occurs. Methods We retrospectively analyzed 999 lung transplant recipients (2007-2024) from the University of Pittsburgh Medical Center. Fully automated deep learning pipelines quantified 61 CT features encompassing lung, vascular, body composition, and calcification compartments from preoperative scans. Clinical and demographic variables were incorporated. CLAD diagnosis and grading followed the International Society for Heart and Lung Transplantation (ISHLT) criteria. Three modeling strategies, including competing risk Fine-Gray (FG), cause-specific Cox (CSC) regression models, and random survival forests (RSF), were used to develop models for predicting CLAD across multiple severity thresholds (any CLAD, Grade ≥2, ≥3, and =4) from 2 to 10 years post-transplant. Model performance was evaluated with 5-fold cross-validation, time-dependent AUCs, and Brier scores. Results The FG model consistently achieved the best discrimination and calibration, with mean AUCs/Brier scores of 0.876/0.124 for any CLAD vs no CLAD, 0.871/0.129 for Grade ≥2, 0.929/0.071 for Grade ≥3, and 0.850/0.142 for Grade=4. The CSC model performed comparably but showed slightly lower calibration, while RSF exhibited stable performance over time but had lower AUCs and higher Brier scores (Figure 1). Key independent predictors included clinical variables (longer ICU stay, ventilatory duration, and greater respiratory support), demographic factors (older age, lower BMI), and vascular image features (increased vascular tortuosity and reduced vessel count). Body composition biomarkers reflecting physiologic frailty were also significant. Higher intramuscular fat density increased risk, whereas greater subcutaneous fat and bone density were protective. Conclusion Quantitative preoperative CT biomarkers provide accurate, grade-specific prediction of CLAD years before spirometric decline. Vascular remodeling and systemic frailty signatures captured through vascular morphology, muscle, and bone metrics are key prognostic indicators. The FG model offered the most discriminative and well-calibrated performance. Integrating these image-derived biomarkers with clinical data offers a practical, noninvasive framework for early CLAD risk stratification, personalized follow-up planning, and timely interventions aimed at improving long-term graft survival. This abstract is funded by: R01HL174570