Primary Graft Dysfunction After Lung Transplantation: A Temporal Classification and Machine Learning Clustering

Primary graft dysfunction (PGD) is a major cause of morbidity and mortality after lung transplantation (LTx). PGD is graded at static time points, limiting insight into its temporal dynamics. Statistical risk-factor analysis may overlook the multifactorial complexity of PGD. Machine learning (ML) may address this but is constrained by small sample sizes. We aim to introduce a temporal PGD classification, perform ML-based clustering and overcome sample-size limitations by generating synthetic patient data.

A prospectively collected database of 794 LTx at University Hospitals Leuven (Belgium) from January 2012 to November 2023 was analyzed. Recipients were classified into 4 temporal PGD phenotypes: no, early, late, and persistent PGD. Unsupervised ensemble k-means clustering was used to group patients based on 30 clinical variables. To generate synthetic data, a Wasserstein Generative Adversarial Network with Gradient Penalty was trained.

Six hundred ninety primary double LTx cases were included. Temporal PGD phenotypes showed significantly different 5-y survival (log-rank test: P < 0.001). K-means clustering revealed stable patient subgroups, with k = 5 and k = 9 identified as optimal cluster numbers. Key features affecting clustering included preoperative intensive care unit stay and postoperative chest x-rays. The synthetic data set preserved the correlation structure of the original data, and the combined data had similar clustering behavior, with optimal clustering at k = 5, k = 7, and k = 9.

This study demonstrates the potential of temporal PGD classification to better understand how the dynamic character of PGD affects survival. ML techniques, including unsupervised clustering and synthetic data generation, could be promising strategies to unravel complex interactions between clinical factors and overcome sample-size limitations.