Introduction

Emergency departments (EDs) are facing unprecedented levels of overcrowding, which leads to increased delays in patient care. By analysing data collected by EDs from patient visits this project aims to develop probabilistic machine learning models to predict patient outcomes (e.g. whether a patient was discharged or admitted to hospital). These models will predict the patient outcome as early as possible in a visit, potentially improving the efficiency of running of EDs and the hospitals they admit patients to.

Explaining the science

Since 2010/11 there has been a 27% increase in emergency admissions to hospital in the United Kingdom, alongside a 10% reduction in the number of inpatient beds. Recently the country has seen some of the worst emergency access performances on record and emergency departments have faced unprecedented levels of overcrowding. Nationally, emergency departments aim to process 95% of patients within 4 hours of their arrival, but this target was last achieved in July 2015, with the most recent national (monthly) average seeing 85% of patients processed within 4 hours. Such delays in care have been linked to increased harm and avoidable mortality and there is, therefore, motivation to process ED patients faster, whilst maintaining the quality of care.

Without the ability to increase the size of an ED or the number of clinical staff, the only way to minimise delays in care is to use the available resources more efficiently. One of the most impactful ways to optimise resource utilisation is to predict the likelihood of hospital admission or discharge of a patient at, or before, the point of arrival to the ED. Accurate predictions will allow more efficient working through several mechanisms; pre-alerting relevant specialties, focused allocation of ED resources, and early downstream bed allocation.

Project aims

The focus of this study will be on evaluating at which points during an ED visit (e.g. at registration or after a given clinical test) a machine leaning model can make accurate predictions of a patient's departure point from the ED. To do this, the project researchers will develop independent models at key points in the ED process, making use of data available at that time.

For example, it's possible to predict the outcome of a patient's visit to the ED at the point of their ED registration using their medical history and any initial observations. Later in the process (as initial clinical diagnostic tests are performed) additional predictions can be made using the results of any clinical diagnostic tests. The developed models can be easily validated as the actual outcome of each patient visit is recorded in a structured way in project partner University Hospitals Southampton's (UHS) ED database, meaning quantitative comparisons can be made.

On completion of the study, the work will have assessed the viability of different machine learning models in supporting ED processes by making point of departure predictions at different points in a patient's ED visit. Results will be directly reported to the ED and, if successful, this pilot project will allow for application of more substantial funding to implement the developed prediction tools into front line clinical practice, and assess impact on health outcomes, clinical decision-making and health care efficiency. 

Applications

The University Hospitals Southampton (UHS) has a Type 1 ED and regional major trauma centre with a consultant-led 24-hour service, which sees over 100,000 patients a year. As part of their process the UHS ED collects a significant amount of real-time, clinical data on arriving patients. The collected data includes descriptions of any co-morbidities, the patients' visible symptoms and any clinical observations. In addition to these, the data may include both near patient (such as blood glucose testing, blood gas analysis and cardiac enzyme testing) and laboratory blood profiles. The UHS ED also makes use of clinical applications to record all relevant information of a patient's visit, such that the collected data can be supplemented by data collected at any previous visits to UHS by a patient.

While the comprehensive nature of the data collected for each patient data makes it a valuable asset to the clinical staff in assessing the patient's medical condition, the breadth of the data means it is infeasible for it all to be reviewed by a clinician. As a NHS Global Digital Exemplar, UHS is looking for novel ways to utilise this data to allow computing and clinical expertise to work in synergy to improve patient flow through the hospital network, and relieve the pressure on the ED by reducing the number of delays in care

Organisers