Predicting emergency department utilization among children with asthma using deep learning models

Abstract

Pediatric asthma is a leading cause of emergency department (ED) utilization, which is expensive and often preventable. Therefore, development of ED utilization predictive models that can accurately predict patients at high-risk of frequent ED use and subsequently steering their treatment pathway towards more personalized interventions, has high clinical utility. In this paper, we investigate the extent to which deep learning models, specifically recurrent neural networks (RNNs), coupled with routinely collected electronic health record (EHR) clinical data can predict the frequency of emergency department utilization among children with asthma. We use retrospective longitudinal EHR data of 87,413 children with asthma aged 0-18 years, who were attributed to one or more healthcare facility for at least 2 consecutive years between 2000-2013. The models were trained for the task of predicting the frequency of emergency department visits in the next 12 months. We compared prediction results of three recurrent neural network (RNN) models: bidirectional long short-term memory (BiLSTM), bidirectional gated recurrent unit (BiGRU), and reverse time attention model (RETAIN), to a baseline multinomial logistic regression model. We assessed the predictive accuracy of the models using receiver operating characteristic curve (AUC-ROC), precision-recall curve (AUC-PR), and F1-score. The results indicated that all RNN models have similar performances reaching AUC-ROC: 0.85, AUC-PR: 0.74, and F1-score: 0.61, compared to AUC-ROC: 0.81, AUC-PR: 0.69, and F1-score: 0.56 for a baseline multinomial logistic regression. Predictive models created from large routinely available EHR data using RNN models can accurately identify children with asthma at high-risk of repeated ED visits, without interacting with the patient or collecting information beyond the patient's EHR.