Machine Learning–Based Prognostic Model for Patients After Lung Transplantation

This prognostic study evaluates a machine learning–based prognostic model for predicting the survival of patients after lung transplantation.

This supplemental material has been provided by the authors to give readers additional information about their work.

Data Collection
A total of 22 characteristics, consisting of 4 recipient factors, 1 donor factor, 4 transplant procedural factors, and 13 posttransplant factors, were collected as follows: age, sex, body mass index (BMI), diagnosis, donor arterial oxygen tension/inspired oxygen fraction (PaO2/FiO2), surgical type, surgical approach, operation time, cold-ischemia time, intensive care unit (ICU) stay, extracorporeal membrane oxygenation (ECMO) type, postoperative ECMO time, preoperative hormone use, grade 3 primary graft dysfunction at 72 h (72 h PGD3), postoperative ventilator time, multidrug-resistant bacterial infection, 6-minute walking test (6MWT), forced expiratory volume at the first second (FEV1), FEV1 percent predicted (FEV1%), forced vital capacity (FVC), FVC percent predicted (FVC%), and FEV1/FVC.Our center prefers a double lung transplant for patients with pulmonary infection and severe pulmonary hypertension.The ECMO type consists of two categories: veno-venous cannulation ECMO and veno-arterial cannulation ECMO.Veno-veno-arterial cannulation ECMO is a variant of venoarterial ECMO and is categorized as such.The maintenance hormone treatment is preoperatively applied to patients according to the original disease condition.MRBI is defined as a bacterial infection that is resistant to multiple antibiotics and causes symptoms.The 6MWT and pulmonary function data were collected from the first examination within 6 months posttransplantly.

Model Development
The grid search method was used for hyperparameter tuning (including number of trees, number of variables to possibly split at each node, and minimum size of terminal node).The bootstrapping resampling method was used to estimate the performance of models fitted by each parameter combination.

Model Validation
The integrated area under the curve (iAUC) of the time-dependent receiver operator characteristic curve (ROC) was used to evaluate the continuous model's discrimination ability.The higher the iAUC of a model, the better its performance in discriminating patients with different outcome statuses.The integrated Brier score (iBS) was applied to estimate the continuous calibration ability of the models.The iBS ranges from 0 to 1, and an iBS of prediction model close to 0 indicates excellent calibration.The iAUC and iBS were estimated from 1 to 48 months at 1-month intervals.Model performance at specific time points (1 month and 1 year) was assessed by the time-dependent area under the curve (tAUC) and prediction error (PE).Meanwhile, the calibration was visualized as a curve between observed survival and predicted survival.A B C D
Additional Information About Data Collection, Model Development, and Model Validation eTable 1. Overall Survival Rate for Patients After Lung Transplantation eTable 2. Predicted Value by RSF Model in Patients With Different Survival Statuses eTable 3. Subgroup Tests for the RSF Model eTable 4. The Performance of RSF Model in 2 Conditions eTable 5.The Performance of Cox Regression Model Based on Stepwise Selection eFigure 1.The Flowchart of Patient Enrollment eFigure 2. Calibration of the Random Survival Forest Model eFigure 3. Consecutive Performance of the RSF and Cox Model eFigure 4. Subgroup Tests for Prognostic Stratification Ability of the Random Survival Forest Model eMethods.

eTable 1. Overall Survival Rate for Patients After Lung Transplantation Survival rates a All patients Conditional on survival to 1 month
a Survival rates were estimated by the Kaplan-Meier method.eTable 2.

eTable 5. The Performance of Cox Regression Model Based on Stepwise Selection Selected variables
The stepwise selection determined 11 factors to develop this Cox regression model.Comparison with the performance of Cox model to RSF model with the same time of prediction.
a b