Machine Learning for Predicting Primary Graft Dysfunction After Lung Transplantation: An Interpretable Model Study.

Area of Science:

• Thoracic surgery
• Transplantation medicine
• Medical artificial intelligence

Background:

• Primary graft dysfunction (PGD) is a critical complication within 72 hours post-lung transplantation (Lung Tx), significantly impacting patient prognosis.
• Accurate prediction of severe PGD (PGD3) is crucial for timely intervention and improved patient outcomes.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting grade 3 PGD (PGD3) after lung transplantation.
• To assess the model's performance in risk stratification for postoperative support.

Main Methods:

• Retrospective analysis of 802 lung transplant recipients (July 2018 - October 2023).
• Identification of independent PGD3 risk factors using logistic regression and LASSO.
• Construction and evaluation of 9 ML models, including Random Forest (RF), for PGD3 prediction.
• Assessment of model calibration, clinical usefulness, and risk stratification capabilities using SHAP values.

Main Results:

• Nine independent clinical risk factors were identified.
• The Random Forest (RF) model demonstrated superior predictive performance across training, internal, and external validation cohorts (AUCs ranging from 0.7975 to 0.9415).
• The RF model showed excellent calibration and clinical utility, and effectively stratified postoperative support needs (ECMO, ventilation, ICU time).

Conclusions:

• The Random Forest model offers optimal performance for predicting PGD3 after lung transplantation.
• The developed RF model is valuable for postoperative risk stratification, aiding in patient management and resource allocation.