Related Experiment Video
Updated: May 6, 2026
Setting Up a Stroke Team Algorithm and Conducting Simulation-based Training in the Emergency Department - A Practical Guide
Published on: January 15, 2017
The Sydney Triage to Admission Risk Tool With Artificial Intelligence (START-AI): Prediction of Inpatient Admission
Michael Dinh1,2, Elizabeth Corbett1,2, Eliot Salmon1,2
1RPA Green Light Institute, Sydney Local Health District, Sydney, New South Wales, Australia.
Objective:
Use artificial intelligence (AI) to extend the Sydney triage to admission risk tool (START) and improve prediction of emergency department (ED) patient disposition.
Methods:
The study was conducted at an inner-city tertiary referral hospital ED. Adult (age ≥ 16 years) presentations from 1 January 2023 to 30 June 2025 were included. Participants were excluded if dead on ED arrival or left ED prior to completing treatment. The primary outcome was admission to an inpatient ward. A sequential ensemble modelling approach was used. To predict patient disposition, the original START was combined (stacked) with vital signs, blood results and CT imaging orders using a gradient boosting decision tree algorithm (XGBoost) and a pre-trained transformer model for clinical free text.
Results:
162,915 cases were analysed with 27.31% overall inpatient admission rate. The final stacked meta-XGBoost model had an area under receiver operating curve (AUROC) of 0.88 (95% CI: 0.88, 0.89) with overall weighted accuracy of 0.84 (95% CI: 0.84, 0.85) and F1 score of 0.83 (95% CI: 0.83, 0.84) in the testing dataset. The model was adequately calibrated with R2 of 0.92 (95% CI: 0.67, 0.99) with a drop-off in correlation at the highest predicted probability ranges (> 0.80). After classifying inpatient stays < 24 h as potential discharges, a sensitivity analysis demonstrated AUROC for the final model of 0.89 (95% CI: 0.88, 0.89).
Conclusions:
An ensemble machine learning model was developed to accurately predict patient disposition from ED using structured and unstructured data. Prototype development and prospective evaluation of START-AI are required to assess model performance in clinical settings.
