Application of Machine Learning to Optimize Management of Children in Hospital with Lower Respiratory Tract Infection

Effective triage can help optimize the use of limited healthcare resources for managing paediatric patients with lower respiratory tract infection (LRTI), the primary cause of death worldwide for under 5 years old children. However, triage decisions do not consider medium to long term needs of hospitalized children. In this study, we aim to leverage data-driven methods using objective measures to predict the type of hospital stay (short or long). We used vital signs (heart rate, oxygen saturation, breathing rate, and temperature) recorded from 12,881 children admitted to paediatric intensive care units in China. We generated multiple features from each vital sign, and then used regularized logistic regression with 10-fold cross validation to test the generalizability of our models. We investigated the minimum number of recording days needed to provide a reliable estimate. We assessed model performance with Area Under the Curve (AUC) using Receiver Operating Characteristic. Our results show that each vital sign independently helps predict hospital stay and the AUC increases further when vital signs are combined. In addition, early prediction of the type of stay of a patient admitted for LRTI using vital signs is possible, even with using only one day of recordings. There is now a need to apply these predictive models to other populations to assess the generalizability of the proposed methods.