How Good Is Machine Learning in Predicting All-Cause 30-Day Hospital Readmission? Evidence From Administrative Data

OBJECTIVES: Hospital readmission is a main cost driver for healthcare systems, but existing works often had poor or moderate predictive results. Although the available information differs in different studies, improving prediction is different from the search for important explanatory variables. With large sample size and abundant information, this study explores state-of-the-art machine-learning algorithms and shows their performance in prediction. METHODS: Using administrative data on 1 631 611 hospital stays from Quebec between 1995 and 2012, we predict the probability of 30-day readmission at hospital admission and discharge. We compare the performance between traditional logistic regression, logistic regression with penalization, and more recent machine-learning algorithms such as random forest, deep learning, and extreme gradient boosting. RESULTS: After a 10-fold cross-validation on the training set (80% of the data), machine learning produced very good results on a separate hold-out test set (20% of the data). The importance of explanatory variables is not the same for different algorithms. The area under receiver operating characteristic curve (AUC) reached above 0.79 at hospital admission and above 0.88 at hospital discharge. Diagnostic codes, which include many different categories, are among the most predictive variables. Logistic regression with penalization also produced good results, but a standard logistic regression failed without penalization. The good results are confirmed by calibration curves. CONCLUSION: Although the identification of those at highest risk of readmission is just 1 step to preventing hospital readmissions, 30-day readmission is highly predictable with machine learning.