Detecting clinical deterioration in respiratory hospital patients using machine learning

Lead Research Organisation: University of Nottingham
Department Name: School of Medicine

Abstract

Background:

There are approximately 6 million emergency admissions to hospitals in England each year. Most people make a good recovery once treatment is started but some patients get worse while they are in hospital. This can for instance be due to catching an infection or developing a blood clot. In the past it was sometimes not noticed that a patient was getting worse until it was too late. Now all hospitals have systems in place to monitor patients on the wards. Nurses regularly check vital signs such as the pulse rate and blood pressure. These are used to calculate an early warning score. If the score is high then the patient may be getting worse and a doctor is called to see the patient.

Early warning scores have saved many lives but they are better at picking up some conditions than others. For instance they are good at picking up severe infections and internal bleeding but sometimes they can fail to pick up breathing problems. Early warning scores also don't work very well in some patients with long-term breathing problems because the score can be high all the time, even when the patient is well. This can result in doctors being called to see patients who don't need any extra treatment. This can distract the doctors from the patients who really need help.


Aims and Objectives:

We aim to develop a better scoring system for monitoring people with breathing problems in hospital. Our goal is that the new scoring system will be better at picking up when people with breathing problems are getting worse and that it will not give too many false alarms.

First we will take a close look at how early warning scores are working in patients with breathing problems. We will check how many times doctors are being called to see this group of patients each day and what extra treatments are being given as a result. This will help us understand what needs to change in our new scoring system.

Next we will gather information about 22,000 patients who were admitted to hospital with breathing problems and examine 1000 cases in detail. We will look at how their vital signs got worse before they developed breathing complications in hospital. We will team up with computer scientists and statisticians to develop a new scoring system which is better than the one we use currently.


Applications and benefits:

Improving the way we monitor patients with breathing problems in hospital will make their care safer. By picking up when things are getting worse at an earlier stage we will be able to prevent severe complications. This is better for patients and will also save money for the NHS because patients will stay in hospital for less time and are less likely to need expensive treatment in Intensive Care.

We will make our results widely available to other researchers and also to the general public. This will allow others to build on our work to further improve the safety of patients in hospital.

Technical Summary

Background and aims:

Early warning scores are composite scores based on routinely measured clinical observations such as pulse, blood pressure and respiratory rate. They are designed to detect deteriorating patients in hospital but currently used scoring systems lack sensitivity and specificity in patients with respiratory conditions. We aim to understand how early warning scores perform in a real-world setting in respiratory patients, and to develop and validate an improved scoring system for this patient group.


Research methods:

Stage 1: We will survey all alerts triggered by a high early warning score in respiratory in-patients at Nottingham University Hospitals NHS Trust over a 3-month period. We will determine through case note review what investigations and interventions were performed as a result of the referrals. This will be used to develop a coding system for common events and interventions.

Stage 2: Clinical observation data will be extracted from hospital systems for the approximately 22,000 admissions to adult respiratory medicine services from 2015-19. 1000 cases in which the emergency team was called at least once will be clinically annotated with reference to the case notes to capture clinically significant events requiring an intervention. The cases will be anonymised and split into a training dataset (750 cases) and a validation dataset (250 cases).

Stage 3: The training dataset will be analysed using machine learning methods including logistic regression, support vector machines and anomaly detection in order to develop novel diagnostic models for detecting clinical deterioration in respiratory patients. The best performing diagnostic models will be tested using the validation dataset and their performance compared with the currently used National Early Warning Score. The area under the receiver operating characteristic curve will be the primary measure of diagnostic accuracy.

Planned Impact

This research will result in an improved scoring system for monitoring patients with respiratory disease in hospital and detecting problems at an early stage. This will result in improved patient safety, fewer adverse events and better clinical outcomes. Early detection of patient deterioration in hospital will reduce the incidence of severe complications requiring expensive treatment in Intensive Care Units. It will also result in quicker patient recovery and reduced length of stay in hospital. This will in turn save money for the NHS. Our novel scoring system will result in fewer false alarms and therefore fewer unnecessary calls to hospital doctors to review patients. This will lead to better use of scarce out-of-hours medical resources so that these can be directed towards patients who are most in need of medical review.

Currently the majority of NHS trusts still record electronic observations and calculate early warning scores using paper charts, which is known to be less reliable than using electronic systems. Developing a successful scoring system which relies on digital technology will encourage the uptake of electronic systems for recording clinical observations and calculating early warning scores throughout the NHS. As well as improving patient safety throughout NHS acute hospitals, this will provide economic benefits to the growing digital healthcare industry in the UK. "Growing the Artificial Intelligence Industry in the UK" is an important priority for the Government and the title of a recent report commissioned by the Business and Culture Secretaries. This project will contribute to developing a pool of people with expertise in the application of data science to medicine and healthcare.

Our public engagement activities will increase the public understanding of the benefits of using digital solutions in healthcare as well as of using anonymised healthcare data in research for the public benefit.

Publications

10 25 50