Machine-Learning for Primary Graft Dysfunction in Lung Transplantation

Purpose The thundering evolution of lung transplantation management during the past ten years and primary graft dysfunction (PGD) new definition have led to new predictive factors of PGD. Therefore, we retrospectively analyzed a monocenter database using a machine-learning method, to determine the predictive factors of grade 3 PGD (PGD3), defined as a PaO2/FiO2 ratio Methods We included all double lung transplantation from 2012 to 2019 and excluded multi-organ transplant, cardiopulmonary bypass, or repeated transplantation during the study period for the same patient. Recipient, donor and intraoperative data were added in a gradient boosting algorithm step-by-step according to standard transplantation stages. Dataset was split randomly as 80% training set and 20% testing set. Relationship between predictive factors and PGD3 was represented as ShHapley Additive exPlanation (SHAP) values. Results A total 478 patients were included in the analysis, 83 (17.3%) had PGD3. Highest performance analysis was achieved at the end-surgery stage (0.87, IC95 [0.867-0.873]) with 6 predictive factors: being under ECMO at some point in the intervention and whatever the reason for its implementation is a predictor; a recipient low total lung capacity is a predictor; ECMO is a predictor whatever the time of implantation; having a cystic fibrosis or a COPD/emphysema is protective while having a lung pulmonary fibrosis or another pathology is predictive. Conclusion Gradient boosting predicted PGD3 with high performance using variables available at the end of double lung transplantation. The implementation of strategies adapted to modifiable variables could make it possible to limit the occurrence of a PGD or its severity.