Automated En Masse Machine Learning Model Generation Shows Comparable Performance as Classic Regression Models for Predicting Delayed Graft Function in Renal Allografts

Background. Several groups have previously developed logistic regression models for predicting delayed graft function (DGF). In this study, we used an automated machine learning (ML) modeling pipeline to generate and optimize DGF prediction models en masse. Methods. Deceased donor renal transplants at our institution from 2010 to 2018 were included. Input data consisted of 21 donor features from United Network for Organ Sharing. A training set composed of similar to 50%/50% split in DGF-positive and DGF-negative cases was used to generate 400 869 models. Each model was based on 1 of 7 ML algorithms (gradient boosting machine, k-nearest neighbor, logistic regression, neural network, naive Bayes, random forest, support vector machine) with various combinations of feature sets and hyperparameter values. Performance of each model was based on a separate secondary test dataset and assessed by common statistical metrics. Results. The best performing models were based on neural network algorithms, with the highest area under the receiver operating characteristic curve of 0.7595. This model used 10 out of the original 21 donor features, including age, height, weight, ethnicity, serum creatinine, blood urea nitrogen, hypertension history, donation after cardiac death status, cause of death, and cold ischemia time. With the same donor data, the highest area under the receiver operating characteristic curve for logistic regression models was 0.7484, using all donor features. Conclusions. Our automated en masse ML modeling approach was able to rapidly generate ML models for DGF prediction. The performance of the ML models was comparable with classic logistic regression models.

Automated summary

In this study, an automated machine learning (ML) modeling pipeline was used to generate and optimize DGF prediction models en masse. The best performing models were based on neural network algorithms, with the highest area under the receiver operating characteristic curve of 0.7595. The performance of the ML models was comparable with classic logistic regression models.