Prediction of the days in hospital for next year using ensemble learning

Abstract

There are many difficulties in predicting problems in real life with an individual model. In order to solve these problems, we use combined these models using an ensemble learning. Combined models can be applied to the prediction of the number of days in hospital. In this paper, we use the 2011 National Sample Cohort Database to construct a combined model using an ensemble learning and predict the number of days in hospital in 2013. Also, we consider the zero inflated poisson models and hurdle poisson models considering that the days in hospital as the response variable have excess zeros. For the problem of overdispersion, we consider the zero inflated negative binomial models and hurdle negative binomial models. As baseline models, we consider the generalized additive models and the generalized linear models using the poisson and negative binomial distributions. Finally, we consider the combined models using ensemble learning. We use the National Sample Cohort Database by National Health Insurance Service and apply eight models to this data. Finally, we suggest the best prediction model by the accuracy criterion of the root mean squared logarithmic error which penalizes an under-predicted estimate greater than an over-predicted estimate.