OBJECTIVES: To evaluate machine learning algorithms in modelling the risk of 30-day heart failure (HF) readmissions in a cohort of commercially insured patients in the US. METHODS: : We used Marketscan commercial claims data (2012-14) to identify a cohort of patients >=18 years admitted with a primary diagnosis of heart failure. Heart failure index admissions and 30 day readmissions were defined using the Centers for Medicare and Medicaid Services (CMS) definitions. Using a combination of CMS defined predictors and empirical analysis, we identified 146 predictors of hospital readmission. Predictors were assessed in the one year period prior to the index heart failure hospitalization. Study data were split into the training set (75%) and the test set (25%). We compared four commonly used machine learning algorithms for binary classification: elastic net regularized generalized linear models, random forests, gradient boosted machines, and support vector machines. We employed 10 fold cross validation on the training set to train each algorithm; predictive performance for each algorithm was assessed using the c-statistic [AUC] on the test set. RESULTS: : In a cohort of 17,631 patients with a qualifying index admission for heart failure, 2,830 patients had a readmission within 30 days. The mean age of the patients in the cohort was 55 years, and 60.2% were males. Based on the c-statistic on the test set, gradient boosted machines performed the best (AUC 0.66), followed by elastic net regularized logistic regression (AUC 0.65), random forests (AUC 0.65), and support vector machines (AUC 0.63). CONCLUSIONS: : Compared to previous models which report an AUC of 0.60, we were able to increase the AUC by 0.06, representing a 12% increase in predictive ability. However, this increase in AUC was largely driven by an increase in the number of predictors rather than any differences in performance between the machine learning algorithms themselves.