This article is more than 5 years old. Information may no longer be current.
Machine-learning algorithms fail to improve prediction of 30-day HF readmissions
Compared with traditional prediction models, researchers determined that the use of machine-learning algorithms did not improve prediction of which patients hospitalized for HF would be readmitted for any cause within 30 days.
Previous models developed to predict 30-day readmission in patients with HF were not discriminatory enough for clinical use, so researchers sought to construct a model using a machine-learning approach and evaluate whether it improved prediction.
The models developed using machine-learning algorithms included a tree-augmented Bayesian network, a random-forest algorithm and a gradient-boosted model. They were compared with traditional statistical methods in a cohort of 56,477 patients from the Get With the Guidelines–Heart Failure registry who were linked to Medicare data.
In the cohort, 70% were assigned to the training set and 30% were assigned to the validation set for the development and testing of model performance.
The outcome of interest was discriminatory capacity as measured by C statistics.
Among the entire cohort, the rate of 30-day rehospitalization was 21.2%, Jarrod D. Frizzell, MD, MS, from University of New Mexico School of Medicine, Albuquerque, and colleagues reported.
C statistics in the validation set for the tree-augmented Bayesian network (0.618), the random-forest algorithm (0.607) and the gradient-boosted model (0.614) were similar to those for a logistic regression model (0.624) and a least absolute shrinkage/selection operator model (0.618), according to the researchers.
A previously validated electronic health records model had a C statistic of 0.589 when applied to the present validation set, they wrote.
“This failure across multiple attempts and methods raises concern that the issue is perhaps less with methodology than with using a subjectively driven outcome or possibly the presence of additional unrecognized covariates of importance,” Frizzell and colleagues wrote. – by Erik Swain
Disclosure : Frizzell reports no relevant financial disclosures. Please see the full study for a list of the other researchers’ relevant financial disclosures.