A11 - Deep Learning-Based Prediction of 10-Year COPD Survival from Paired Inhale-Exhale Computed Tomography Scans

Introduction: : Chronic obstructive pulmonary disease (COPD) is a leading cause of death globally. Existing prognostic and mortality prediction tools often fail to capture the full heterogeneity of the disease. Traditional models like the BODE index rely on generic, rule-based clinical features that overlook subtle biomarkers of COPD progression in the lungs. While chest CT imaging offers detailed anatomical and functional insights, it remains underutilized, with most studies focusing on a limited set of features such as emphysema, bronchial wall thickening, and air trapping. We hypothesize that 3D deep learning models can learn richer anatomical and functional biomarkers directly from the data. This work introduces a novel approach that uses paired inhale-exhale CT scans to predict 10-year mortality in COPD patients.

Materials and

Methods: : We used the COPDGene dataset, a large cohort (N ≈ 10,000) with rich longitudinal follow-up data. The dataset was split into training (70%), validation (15%), and testing (15%) using stratified sampling to maintain mortality event balance. We finetuned MedicalNet, a 3D ResNet pretrained on medical image segmentation, separately on inhale and exhale CT volumes to predict binary mortality (called INSP-Net and EXP-Net, respectively). From each model, we extracted a 512-dimensional feature representation. These inhale and exhale embeddings were then fused using a transformer-based architecture with attention pooling. An additive hazard layer was applied to predict year-by-year mortality risk over a 10-year horizon. Model performance was evaluated using the concordance index (C-index) and time-dependent area under the curve (AUC). As a clinical baseline, we computed the BODE index C-index on the same test cohort.

Results, Conclusions, and Discussions:: Results and Discussion

Our results show that deep learning applied to 3D CTs alone can effectively predict 10-year mortality in COPD patients, outperforming traditional clinical models. The inhale-only model (INSP-Net) achieved a test AUC of 0.76 and accuracy of 80.27%, while the exhale-only model (EXP-Net) reached an AUC of 0.77 and accuracy of 79.97%. These strong single-view baselines demonstrate that both inhale and exhale CTs contain meaningful prognostic information.
The fused inhale-exhale model, which leverages a transformer-based architecture with attention pooling and an additive hazard layer, achieved a C-index of 0.75 and a mean time-dependent AUC of 0.77. Per-year AUCs (years 1–10) were as follows: [0.69, 0.78, 0.79, 0.79, 0.77, 0.78, 0.77, 0.77, 0.77, 0.78]

Our model consistently outperformed the BODE index baseline (C-index: 0.71) and shows potential for improving upon the single-view imaging models, highlighting the added value of functional information from paired inhale-exhale scans. We observed a dip in AUC at year 1 (0.69), which we hypothesize may be due to the challenge of stratifying short-term mortality from imaging alone. Performance improved and remained stable through years 2–10 (ranging from 0.77 to 0.79), suggesting the model effectively captures both medium- and long-term risk.

We plan an ablation study to isolate the impact of the feature fusion, transformer, attention pooling, and hazard modeling. While this work focused on imaging alone, future extensions will explore fusing clinical features alongside the learned imaging representations to improve generalizability and robustness.

Conclusion

This study demonstrates the effectiveness of 3D deep learning models trained on inhale-exhale CT for long-term mortality prediction in COPD. The imaging-only model surpassed the BODE index and achieved strong time-dependent AUCs. The transformer-based fusion approach captures complementary lung function and anatomy. Ongoing work focuses on ablations and adding clinical variables to improve risk stratification and clinical utility.

Acknowledgements and/or References (Optional)::