Developing a Decision Support System at Emergency Room triage (DESSERT) for predicting health outcomes

A decision support system at emergency room triage. No predefined intervention(s)/exposure. We will access patient medical records with no active involvement from participant required. All medical records of adult patients presenting at Waitemata District Health Board (WDHB) hospitals’ emergency departments (ED) in Auckland, New Zealand during the 5-year study period from 2016 to 2021 will be included in this retrospective cohort study. The duration of follow-up of each participant is up to 6 months post-index ED presentation. The first four years’ retrospective data of this cohort (from July 2016 to June 2020) will be used to develop a Decision Support System at the time of ED triage for predicting hospital admission and long ED length of stay. The prospectively collected cohort data from July 2020 to June 2021 will be used to validate the performance of this Decision Support System. The actual period of the development and validation cohorts may now vary due to the general health recommendations during the global pandemic. We will use two methods to develop a Decision Support System (prediction model). One is applying machine learning techniques to develop a Decision Support System (prediction model) with the highest predictive ability for hospital admission and longer ED length of stay. The following prediction algorithms will be investigated in this study, logistic regression, support vector machines, Naive Bayes algorithm, decision trees, random forest, gradient boosting and deep learning. In addition, we will also use the traditional logistic regression to develop this Decision Support System for predicting health outcomes. To ensure high statistical and clinical significance, only variables with a p value of <0.05 will be included in multivariable logistic regression analyses to form the proposed Decision Support System. The modelling performance of different prediction models will be assessed by receiving operator characteristic (ROC) curve with a bootstrapping method using 10,000 replicates to calculate 95% confidence intervals, to assess the predictive performance for predicting health outcomes at the time of ED triage. Our research group involved experienced triage nurses, ED clinicians, Geriatrician, epidemiologists and biostatisticians. We will have regular meetings (1-2 hours every 2 months for 24 months) to discuss the study progress and findings. There was no further planned surveys/focus groups/interviews or other interactions with staff or patients. However, we will seek advice, when needed, from Emergency Medicine Specialist, Operations Manager Emergency Department, ED clinicians, Triage nurses and other researchers.