Early detection of at-risk patients of re-presentation to emergency department using a recurrent neural network (Preprint)

BACKGROUND Re-presentations to emergency departments (EDs) have been directly associated with increased healthcare cost and length of stay, poorer quality of care and increased morbidity and mortality. Early detection of at-risk patients to EDs can reduce unnecessary re-presentations and provide provision of better quality of care and healthcare planning. Conventional risk predictive models, however, have difficulties when the at-risk patients have diverse and complex disease states or demographic profiles. These models also ignore related temporal patient information such as changes in their disease state and personal circumstance which can be used to model the progression of risks. OBJECTIVE Our aim is to develop a temporal risk predictive model based on recurrent neural network (RNN) can understand temporal relationships between different times of patient presentations to EDs and improve the predictive modelling. METHODS We used the data extracted from Health Information Exchange (HIE) system, which included all available ED records from the Nepean hospital in Australia from the period 1 January 2009 to 30 June 2016. A total of 343,014 ED presentations were identified from 170,134 individual patients. We used the variables including age, marital status, indigenous status, mode of arrival, mode of separations, referred to on departure and diagnosis code which have shown to be correlated to frequent presenters to EDs. We evaluated our RNN model by comparing it to other conventional predictive models using the area under to receiver operating characteristics curve (AUROC). All models were trained using the ED data extracted from the 6 to 12-months period by setting an interval that is divided into an observation window and a prediction window. We further proposed a context-based patient representation learning (CPRL) framework to better characterise the feature representation of patient data and discussed the general extension of our CPRL framework as an optimisation algorithm to improve the feature representation of patient data. RESULTS Using a 9-month observation with 1-month prediction window (i.e., prediction of at-risk patients of re-presentation to ED in next 1-month), the AUROC for the RNN model was 71.60% compared to AUROCs for logistic regression (57.18%), Naves Bayes (56.35%) and random forest (56.02%). The at-risk patients presented to the ED more frequently (i.e., time (day) differences between presentations become shorter) when their marital status was changed (e.g., from ‘Married’ to ‘Separated’ or ‘Separated’ to ‘Divorced’). These patients also consistently had similar diagnoses during the observation period, indicating that these groups of patients may be the focus of certain integrated cares / interventions to improve the quality of care and reduce the unnecessary re-presentations. CONCLUSIONS Our findings indicate that our RNN improves the predictive modelling, is robust and can effectively understand the disease state and personal circumstance changes within patients over time. We suggest that our model highlights the gaps in ED interventions and can be used to develop tailored integrated cares / interventions.