Neural network-based multi-task learning for inpatient flow classification and length of stay prediction

Abstract

Inpatient unit resources are among the most expensive and valuable resources for healthcare organizations. Inpatient resources such as room, bed, and medical devices can be more efficiently managed if we can predict inpatient flow and length of stay (LOS) before admission and inpatient bed assignment [1]. Patient LOS prediction has been researched individually using classical machine learning methods, such as linear regression, regression trees, random forest, and neural networks for a long time. Inpatient LOS and flow share many common features in training predictive models because both are closely related to relevant features such as recovery status and surgery types. Besides, these two tasks are closely related. For example, a patient with a more complex inpatient flow tends to have a longer LOS. This paper is the first comprehensive study that links them together as multi-tasks and develops an artificial neural network-based multi-task learning model (ANNML) for mixed types of task prediction in inpatient LOS and flow identification. The constructed multi-task learning model was tested on a real-life dataset collected from a large hospital in New York City and compared with four single-task learning models. The results show that ANNML can use the most relevant features to achieve a better prediction accuracy for both task types and has less overfitting and testing variance than single-task learning models.

Lu He 何璐

Faculty of Supply Chain Management

My research interests include systematic resource optimization, multitask prediction, and predictive-driven mixed integer programming.