Comparison of machine-learning and logistic regression models to predict 30-day unplanned readmission: a development and validation study

We compared the predictive performance of gradient-boosted decision tree (GBDT), random forest (RF), deep neural network (DNN), and logistic regression (LR) with the least absolute shrinkage and selection operator (LASSO) for 30-day unplanned readmission, according to the number of predictor variables and presence/absence of blood-test results. We used electronic health records of patients discharged alive from 38 hospitals in 2015–2017 for derivation (n=339,513) and in 2018 for validation (n=118,074), including basic characteristics (age, sex, admission diagnosis category, number of hospitalizations in the past year, discharge location), diagnosis, surgery, procedure, and drug codes, and blood-test results. We created six patterns of datasets having different numbers of binary variables (that ≥5% or ≥1% of patients or ≥10 patients had) with and without blood-test results. For the dataset with the smallest number of variables (102), the c-statistic was highest for GBDT (0.740), followed by RF (0.734), LR-LASSO (0.720), and DNN (0.664). For the dataset with the largest number of variables (1543), the c-statistic was highest for GBDT (0.764), followed by LR-LASSO (0.755), RF (0.751), and DNN (0.720). We found that GBDT generally outperformed LR-LASSO, but the difference became smaller when the number of variables was increased and blood-test results were used.