Proposed Topic (Most preferred): :
Research and Innovations (new projects / technology / innovations / service models)
Authors (including presenting author) :
Fong KM (1), Ng GWY (1)(2), Mak KK (4), Cheung HL (4), Cheng W (3), Ng T (3), Lau A (3), Lam A (3), Wong K (3), Yeung T (3), Mak C (5)
Affiliation :
(1) Intensive Care Unit, Queen Elizabeth Hospital
(2) Multi-disciplinary Simulation and Skills Centre, Queen Elizabeth Hospital
(3) Information Technology, Kowloon Central Cluster
(4) Cluster Nursing Division, Kowloon Central Cluster
(5) Department of Neurosurgery, Queen Elizabeth Hospital
Introduction :
Early warning scores, such as Modified Early Warning Scores, have been widely adopted in general wards for decades to identify high-risk patients. This assists healthcare providers in providing relevant monitoring and timely intervention to patients before clinical deterioration, potentially improving patient outcomes. With the increasing accessibility of electronic health records, it is now feasible to utilize various laboratory parameters to automatically predict patient deterioration.
Objectives :
The objective of this project is to develop a machine learning prediction model that can forecast clinical deterioration within the next 72 hours.
Methodology :
Patients admitted to 9 medical and neurosurgical wards at Queen Elizabeth Hospital (QEH) from 1 Jan 2022 to 30 Nov 2022 were recruited for this study. Data, including patient demographics (such as age and gender) and the 15 most recent laboratory results (e.g., hemoglobin, creatinine, alanine transaminase), were collected from CDARS. Each record indicated whether the patient required resuscitation within the next 72 hours. The initial dataset comprised 86,000 rows. A logistic regression model was developed and deployed to the QEH Doctor Dashboard as a production pilot in medical wards from June to July 2023. The model was then validated using prospectively collected data during the production pilot.
Result & Outcome :
The initial model achieved an area under the receiver-operating characteristics curve (AUC) of 0.63, an accuracy (ACC) of 0.77, a true positive rate (TPR) of 0.45, and a true negative rate (TNR) of 0.76. Subsequently, the training dataset was expanded to include all general wards in QEH, and the prediction outcome was changed to a composite outcome of the need for resuscitation or death within the next 72 hours. The expanded dataset (475,000 rows) and the use of the composite outcome resulted in an improvement in prediction performance. The latest model achieved an AUC of 0.85, an ACC of 0.75, a TPR of 0.80, and a TNR of 0.75.
The machine learning model developed for predicting patient deterioration demonstrated satisfactory performance. Currently, efforts are underway to enhance the model by incorporating patients' vital signs, collecting ongoing data, and fine-tuning the prediction outcomes. Integrating the model with the Doctor Dashboard may play a role in supporting clinical decision-making and improving patient outcomes.
