A de-identified dataset was extracted from the Electric Health Record (EHR) system at the Royal Melbourne Hospital, Melbourne, Australia. This dataset encompassed all patient visits to the Emergency Department (ED) between June 2019 and December 2022, amounting to 341,026 ED visits and involving 202,520 unique patients.
The aim of this study was to develop ML models capable of predicting ED visit outcomes for adult patients at different time points. Two primary binary outcomes were defined, each accompanied by a corresponding secondary ternary outcome. The first primary outcome determined whether the Length of Stay (LOS) exceeded 4 hours, encompassing both the ED and the subsequent hospital stay. The related secondary outcome categorized LOS into: (1) within 4 hours; (2) between 4 and 24 hours; (3) exceeding 24 hours. The other primary outcome determined the discharge decision (DD) for patients, whether they were discharged or required continued stay. The corresponding secondary outcome indicated whether the patient was: (1) discharged; (2) admitted to short stay; (3) admitted to an inpatient bed. Three time points for predictions were selected: 10, 60, and 120 minutes, commencing from the end of triage.
The main focus of the talk will be to discuss the many practical challenges of working with such large medical real-world data sets and how we overcame them. We will present interesting conclusions for future work based the overall results of the ML models, which are approx. 0.90 AUC.
Professor Aickelin's (Head of School of Computing and Information Systems) expertise is in Artificial Intelligence and Data Mining for Medicine. Prior, he was Vice-Provost of the University of Nottingham Ningbo China and a Strategic Adviser for Artificial Intelligence to the UK Research Councils and Government.