Predicting Discard of Both Kidneys From Deceased Donors Using Machine Learning: Model Development and Comparison Study

Supplementary material for the study “Predicting Discard of Both Kidneys From Deceased Donors Using Machine Learning: Model Development and Comparison Study”. The study compares logistic regression (LR), a decision tree (DT), a random forest (RF), extreme gradient boosting (XGB), a multilayer perceptron (MLP) and an ensemble of LR, RF, XGB and MLP (DE) for predicting whether at least one kidney of a deceased donor is transplanted. The donor data of the German Organ Procurement Organization (DSO) cannot be shared for data protection reasons. This archive contains the results derived from them that are needed to reproduce the evaluation without access to the data: <MODEL>_predictions.csv (all six models): predictions for the test set (n = 817 donors). Prob_0 and Prob_1 are the predicted probabilities of class 0 (both kidneys discarded) and class 1 (at least one kidney transplanted). Prediction is the predicted class at a threshold of 0.5, and Targets is the observed outcome. <MODEL>_val_predictions.csv: the same for the validation split (n = 490), which was used for threshold selection and calibration. <MODEL>_feature_selection_fs_df.csv (all models except DE): all 500 trials of the feature selection (NSGA-II, 3-fold cross-validation on the training set). Each params_<feature> column indicates whether a candidate feature was included in a trial, so the column names list all 1,091 candidate features. The table also gives the objective (−MCC + 0.0005 × number of features), the cross-validated MCC, the number of features and the generation of each trial. <MODEL>_trial_data.csv (all models except DE): the hyperparameter search space and all 100 trials of the hyperparameter optimization (TPE, 5-fold cross-validation on the training set), with the hyperparameters and the cross-validated MCC of each trial. requirements.txt and python-version.txt: exact package versions and Python version of the analysis environment. The analysis code is available at https://github.com/PeerHan/donor-kidney-discard-prediction (version v1.0-revision).