Prediction of Pulmonary Graft Dysfunction After Double-lung Transplantation (PGD3-AI Study)

The standardized anesthetic management has been previously described 18 and is detailed on the web site http://anesthesie-foch.org/protocoles-anesthesie/ ("The Foch lung transplant anesthesia protocol").

Continuous variables are presented as median + interquartile range (IQR) or mean and 95%CI, and were compared using independent T-test or Mann-Whitney test. Categorical variables are presented as n (%) and were compared using Chi-squared test or Fisher's exact test. We applied machine learning algorithm to predict 3-day ahead primary graft dysfunction after lung transplant surgery among patients. Machine learning is a branch of artificial intelligence where computer systems can learn from available data and identify patterns with minimal human intervention. Machine learning algorithm tests on data and performance metrics were used to obtain the higher performing algorithm. In this study, we performed a XGBoost (Gradient Boosting) algorithm which was a combination of decisions trees. Each decision tree typically learned from its precursor and passed on the improved function to the following. The weighted combination of these trees provided the prediction.

No particular data transformation has been performed on numerical variables. Categorical variables have been encoded as integer, without any further pre-processing steps. In particular, no specific processing has been performed to deal with missing data. The default behavior of XGBoost has been used. It consists in treating missing data as a specific modality. During the training step of XGBoost models, missing values are treated as other values, and left or right decisions at any branch of a tree are learned by optimizing the outcome.

In order to reflect the sequential nature of this predictive medicine problem, nine steps have been defined to take into account incrementally observed variables acquired at various stages of the surgery.

Step 1: recipient variables Step 2: donor variables Step 3: arrival in the OR Step 4: after anesthetic induction Step 5: during first pneumonectomy Step 6: after first graft implantation Step 7: second pneumonectomy Step 8: second graft implantation Step 9: end surgery status At each of the nine steps, a cross-validation procedure is employed to assess the predictive performance of a machine learning model (XGBoost). One repetition of the cross-validation procedure is designed as follows: the dataset of subjects is randomly split into eight disjoint parts. Successively, the performance of the XGBoost model on each of the eight subset,while training the machine learning model using the remaining seven subsets. For such a repetition, the predictive probability of 3-day ahead primary graft dysfunction for each subject is retained to finally compute the area under ROC (receiving operator curve). To evaluate the variability of the predictive performance of the machine learning model, this cross-validation procedure is repeated fifty times, with randomly chosen subjects partitions. For each of the fifty times eight times nine (repetitions, partitions, surgical steps), hence 3600 models training, a conservative approach has been adopted for XGBoost training, consisting in a unique set of training parameters. These parameters have been chosen to prevent overfitting due to a relatively small number of subjects compared to the number of variables, especially categorical variables, which yield a high degree of freedom. Specifically XGBoost has been trained for 400 rounds (no early stopping), a maximum depth of 5 for each tree, a minimum child weight of 3, and a learning parameter eta equals to 0.0002. Besides those conservative parameters chosen to prevent overfitting, only 40 percents of available columns are selected for tree construction at each round, and 95 % of subjects. These parameters have been kept fixed and chosen to ensure stability of results. Small perturbations around these values could result in local performance improvements, but would not be practically chosen given the size of the dataset.

In order to gain some insights into the most useful variables in terms of predictive power, we then conducted a post-hoc analysis based on the following methodology: at each surgical step, 400 models have been trained for the repeated cross-validation procedure. For each model, we retain the rank of each variable as given by the variable importance procedure of XGBoost. The average rank of each variable for each step is then computed by averaging the ranks obtained by variables for each of the 400 models. At step 9, variables are ordered based on their average rank (increasing average ranks). They are then incrementally used as input of a new cross-validation procedure (repeated 20 times).