Abstract:
|
Predicting hospital readmission can reduce health care costs and improve patient satisfaction. With patient longitudinal admission history, a convenient approach to estimate a predictive model and evaluate its accuracy is to use each patient’s first hospitalization only. This avoids the complexity of modeling series of hospitalizations but potentially incurs a loss of information. We propose to estimate prediction models using the generalized linear mixed model (GLMM), typically a logistic model for the probability of a 30-day readmission. To create predictions on future subjects, we derive the posterior density of patient-specific random effects considering all available data to date. Specifically, given the patient-specific posterior densities, we use importance sampling to estimate the posterior mean of the random effects, and create predicted values using the estimated GLMM parameters and posterior mean. We apply our method to 2016–2019 Medicaid hospitalization claims data on patients with diabetes. The results show that our method robustly identifies hospitalized patients with diabetes who are at high risk of readmission.
|