PREDICT-DL: Performance of Paediatric Real-time Emergency Department admission and Inpatient Care Time prediction with Deep Learning

Patient flow prediction modelling with machine learning using triage data. Machine learning will use a ensemble model of XGboost, multilayer perceptron and TabNET with a logistic regression blend and natural language processing to interpret and embed the triage free text. Outcomes of patient flow predicted are admission from ED, with short stay vs medical vs surgical admissions to also be predicted, as well as the length of stay. The study will use the previous 5 years of data for training and simulation prior to the intervention, with the intervention then using 6 months of live prospective data. During this time, the bed managers who receive the model outputs will be involved in qualitative interviews as an additional outcome and source of data. Bed manages will receive hourly predictions (with confidence intervals) on a digital bed occupancy dashboard and they will receive training in how these predictions are created and what they mean via educational sessions and a reference guide available at the time of use. This dashboard is already in clinical use, it just does not provide any ahead of time predictions. It is intended that this can help inform their capacity planning, though resource allocation ahead of the time of actual admission. This will given them information from the time of a patients triage, giving them valuable lead time, but the system will not suggest how to change capacity planning, it will only provide predictions of what the bed demand will be in the upcoming hours. There is no quantitative method of assessing adherence or click through based on system limitations, though assessment will occur via qualitative interviews both early and late in the intervention.