B44-32 Synergistic Multimodal Deep Learning for Survival Prediction in COPD

Rationale Chronic obstructive pulmonary disease (COPD), the fourth leading cause of death globally, exhibits marked heterogeneity in its structural, functional, and clinical manifestations. This makes individualized prognosis difficult. Accurate survival prediction is clinically critical for guiding treatment intensity, follow-up, and advanced care decisions. Conventional indices such as GOLD stage and FEV1 reflect airflow limitation but overlook structural disease burden and progression risk. In contrast, chest CT captures emphysema, airway remodeling, and vascular changes that often precede functional decline. This study aims to determine whether integrating CT-derived image features with clinical variables through deep learning could yield more precise, patient-specific survival modeling for COPD. Methods A total of 757 patients (1,534 chest CT scans) from the COPD Specialized Center for Clinically Oriented Research (SCCOR) cohort at University of Pittsburgh were analyzed, characterized by variables including demographics and spirometry. A synergistic deep learning framework was developed, comprising (1) a 3D ResNet-18 image encoder to extract spatial representations, (2) a multimodal fusion layer with projection and interaction submodules aligning and concatenating imaging and clinical representations into a unified latent space, and (3) a DeepHit survival head modeling discrete-time event probabilities for non-proportional risk learning. Models were trained and evaluated under three configurations, clinical-only, CT-only, and multimodal, using five-fold cross-validation. Performance was evaluated using the concordance index (C-index), time-dependent AUC, and Brier score, with significance assessed by paired two-tailed t-tests. Results The multimodal model consistently outperformed all single-modality baselines. In cross-validation, it achieved the highest mean C-index of 0.804 (95% CI: 0.791-0.817), significantly exceeding the performance of the clinical-only model (0.780, 95% CI: 0.756-0.805; p = 0.025) and the CT-only model (0.779, 95% CI: 0.763-0.796; p = 0.004) (Figure). Although the multimodal model also achieved the lowest Brier score (0.108, 95% CI: 0.095-0.121) compared with the clinical (0.112, 95% CI: 0.101-0.122) and CT (0.111, 95% CI: 0.102-0.121) models, these differences were not statistically significant. Time-dependent AUCs for the multimodal model were 0.825, 0.843, and 0.815 at the 3-, 5-, and 10-year horizons, respectively, indicating superior and more stable predictive performance over extended follow-up periods. Conclusions Integrating deep-learned CT features with clinical variables improves survival prediction for COPD and enables more accurate, patient-specific risk estimation. CT-derived structural patterns provide complementary information on disease burden and progression beyond spirometry. This multimodal framework offers a generalizable foundation for precision prognosis and longitudinal monitoring, supporting early identification of high-risk patients, personalized follow-up, and AI-assisted clinical decision-making in COPD. This abstract is funded by: R01HL174570 and R01CA237277