Prognostic models for COVID-19 to support risk stratification in secondary care

Lead Research Organisation: University of Birmingham
Department Name: Institute of Applied Health Research

Abstract

As of mid-July 2020, almost 600,000 people have died with COVID-19 (coronavirus) worldwide. Some patients who are admitted to hospital with COVID-19 experience a rapid worsening of their symptoms and go on to need intensive care treatment, ventilation (to help them breathe) or die. Because the virus is new and affects different people in different ways, doctors find that their clinical experience is not enough to help them to predict which patients are most likely to develop severe symptoms or die, and there is no tool which can help them to do this.

Therefore, the aim of our study is to develop tools that will help healthcare professionals to identify patients at high risk of needing intensive care treatment or ventilation, or of dying, as well as patients at low risk who can be safely discharged from hospital. This may provide an early opportunity to treat patients at high risk, while also making best use of limited hospital resources. We will do this by using anonymised patient data from hospitals in the UK to:

1. Develop a model which uses patients' symptoms, test results, and other information to predict their risk of needing intensive care treatment/ventilation, or dying.

2. Find groups of patients with similar test results and explore how their condition progresses.

Technical Summary

Aim: The overarching aim of this study is to develop tools to aid clinicians to appropriately risk stratify confirmed COVID-19 cases in a UK hospital setting. There are 2 objectives:
1. Develop and externally validate prognostic models for i) the composite outcome of ITU admission and/or mechanical ventilation, and ii) death in a UK secondary care setting.
2. Identify distinct clusters based on biomarkers in patients diagnosed with COVID-19 and map prognostic trajectories in these patient groups, particularly with respect to progression to requiring ITU admission, ventilation or death.

Design: Retrospective cohort analyses using routinely collected secondary care data. For both the prognostic models and cluster analysis participants will all be followed from index (COVID-19 test) date until the earliest of outcome date or study end (latest available data). Patients will be censored 30 days after index date.

Study population: Hospitalised patients of all ages diagnosed with COVID-19 (defined as a positive test result from one or more RT-PCR test).

Primary outcomes: Prognostic models: i) A composite outcome of ITU admission and/or mechanical ventilation; and ii) death. Cluster analysis: Clinically useful clusters of indicators and associated prognosis: probability of ITU admission, mechanical ventilation or death. We will also explore whether clustering gives valuable information on potential clinical course of the illness, for example if it resulted in myocardial involvement, thromboembolic events, gastrointestinal symptoms and severe pneumonia.

Candidate predictors for the prognostic model: Predictors will be explored based on availability of data and existing evidence/biological plausibility. These will include: demographic characteristics; symptoms; frailty; vitals; biomarkers; radiography; physiological tests; and medications/treatments.
 
Description The aim of our study was to develop tools that will help healthcare professionals to identify patients admitted to hospital with COVID-19 who are at high risk of needing intensive care treatment or of dying, as well as patients at low risk who can be safely discharged from hospital. A tool such as this may provide an early opportunity to treat patients at high risk, while also making best use of limited hospital resources. We have done this by using anonymised patient data to:

1. Develop prediction models which use patients' symptoms, test results, and other information to predict their risk of dying or needing intensive care treatment using data from patients with COVID-19 admitted to University Hospitals Birmingham (UHB). We have also checked how well these models work in a different group of patients.

2. Find groups (clusters) of patients with similar test results and explore how their condition progresses.

To develop the prediction models, we used data from more than a thousand patients admitted to hospital with COVID-19, of whom more than a quarter died within 28 days of being admitted to hospital, and nearly one in five required intensive care treatment. Factors that were found to predict death or intensive care admission included patient characteristics like age, symptoms such as breathlessness and fever, measurements such as blood pressure and oxygen level, laboratory test results such as white blood cell count and albumin level, and scores such as the Glasgow Coma Scale (which measures consciousness). The models were very good at differentiating between patients who would go on to develop one of these outcomes and those who would not, in both the UHB patients and patients from other hospitals.

While we were carrying out our study, another score called the 4C score was published which also aims to predict which hospitalised patients with COVID-19 are most likely to die. Therefore we also tested how well this score worked in our patient dataset. We found that this score and our new model were both good at predicting which patients would go on to die. However, the 4C score uses information on comorbidities (diseases) that patients have, which are not always well recorded in the patient record at the time the patient is admitted to hospital; our models did not use comorbidities.

Finally, we have also found that there are distinct groups or clusters of patients based on their symptoms and test results, and that different clusters have a different likelihood of dying or being admitted to intensive care.
Exploitation Route While we found that the prediction models we developed performed well, our models were developed using data from the first wave of COVID-19 infections, and the way COVID-19 affects individuals has altered following the successful roll-out of vaccines. It would therefore be important to validate (and, if necessary, update) the models using more recent data before using them in a healthcare setting.

Nevertheless, the methods we used to develop the prediction models and to conduct the cluster analysis may be useful to researchers in future studies attempting to risk stratify or identify meaningful patient subgroups.
Sectors Healthcare
URL https://bmjopen.bmj.com/content/12/1/e049506
 
Title Risk calculators for mortality and ITU admission for patients admitted to hospital with COVID-19 
Description The web app contains two calculators: one for calculating risk of death within 28 days of admission for patients admitted to hospital with COVID-19, and a second for calculating risk of intensive therapy unit (ITU) admission for patients admitted to hospital with COVID-19. These calculate % risk of death/ITU admission using the coefficients from the prognostic models we developed as part of the grant (the prognostic model coefficients are available at https://doi.org/10.1101/2021.01.25.21249942). 
Type Of Technology Webtool/Application 
Year Produced 2021 
Impact N/A 
URL https://uhb-calc.netlify.app/