Developing approaches to incorporate donor‑lung CT images into ML models to predict severe PGD after lung transplantation
We present a multi‑modal machine learning framework integrating donor‑lung CT imaging and clinical data to predict severe (grade 3) primary graft dysfunction following bilateral lung transplantation. In a retrospective cohort of 160 recipients treated over a 10‑year period, four computer‑vision strategies were assessed; a 3D ResNet model achieved AUROC/AUPRC of 0.63/0.48 (median). Using a late‑fusion approach combining CT features and EHR data improved median AUROC to 0.74 and AUPRC to 0.61, outperforming imaging‑only or clinical‑only models.
This work supports the added value of CT‑derived imaging phenotypes in donor acceptance decision tools and paves the way toward a clinically deployable decision support system.