News & Events

New Early warning system forecasts who needs critical care for Covid-19

17.05.2022

An algorithm predicts which hospitalized patients have the highest risk of death and need intensive care treatment

 

An algorithm, developed at the University of Vienna and the Max Planck Institute for Multidisciplinary Natural Sciences, can predict which hospitalized patients have the highest risk of death and require intensive care treatment, regardless of immune protection status and viral variant. The tool, developed by an international team led by David Gómez-Varela, uses artificial intelligence (AI).

The algorithm, with the name of "COVID-19 Disease Outcome Predictor" (CODOP), could help physicians with directing intensive care resources to those who need them most, and is especially valuable for countries with limited resources.

“The appearance of new SARS-CoV-2 variants, waning immune protection and relaxation of mitigation measures, means we are likely to continue seeing surges of infections and hospitalizations” explains the leader of this international project and senior author David Gómez-Varela, former Max Planck Group Leader and current Senior Scientist at the Division of Pharmacology and Toxicology, University of Vienna, Austria.  “There is a need for clinically valuable and generalisable triage tools to assist the allocation of hospital resources for Covid-19, particularly in places where resources are scarce. But these tools need to be able to cope with the ever-changing scenario of a global pandemic and need to be easy to implement.”

To develop such a tool, the scientists used biochemical data from routine blood draws performed in nearly 30,000 patients hospitalised in over 150 hospitals in Spain, the US, Honduras, Bolivia and Argentina between March 2020 and February 2022. This means they were able to capture data from people with different immune statuses – vaccinated, unvaccinated and those with natural immunity – and from people infected with every SARS-CoV-2 variant, from the virus that emerged in Wuhan, China, to the latest Omicron variant. “The intrinsic variability in such a diverse dataset is a great challenge for AI-based prediction models” says lead author Riku Klén, Associate Professor at the University of Turku, Finland.

The resulting algorithm uses measurements of 12 blood molecules that are normally collected during admission. This means the predictive tool can be easily integrated into clinical care of any hospital.

CODOP was developed in a multistep process, initially using data from patients hospitalised in more than 120 hospitals in Spain, to ‘train’ the AI system to predict hallmarks of a poor prognosis. The next step was to ensure the tool worked regardless of patients’ immune status or Covid-19 variant, so they tested the algorithm in several subgroups of geographically dispersed patients, either unvaccinated or vaccinated, and infected with different SARS-CoV-2 variants. The tool still performed well at predicting the risk of in-hospital death during this fluctuating scenario of the pandemic, suggesting the measurements CODOP is based on are truly meaningful biomarkers of whether a patient with Covid-19 is likely to deteriorate.

To test whether the time of taking blood tests affects the tool’s performance, the team compared data from different time points of blood drawn before patients either recovered or died. They found that the algorithm can predict the survival or death of hospitalised patients with high accuracy until nine days before either outcome occurs.

Finally, they created two different versions of the tool for use in scenarios where healthcare resources are either operating normally or are under severe pressure. Under normal operational burden, doctors may opt to use an ‘overtriage’ version, which is highly sensitive at picking up people at increased risk of death at the expense of detecting some people who did not require critical care. The alternative ‘undertriage’ model minimizes the possibility to wrongly select people at lower risk of dying, providing doctors with greater certainty that they are directing care to those at the highest risk when resources are severely limited.

“The performance of CODOP in diverse and geographically dispersed patient groups and the ease of use suggest it could be a valuable tool in the clinic, especially in resource-limited countries,” remarks David Gómez-Varela. “We are now working on a follow-up dual model tailored to the current pandemic scenario of increasing infections and cumulative immune protection, which will predict the need for hospitalisation within 24 hours for patients within primary care, and intensive care admission within 48 hours for those already hospitalised. We hope to help healthcare systems restore previous standards of routine care before the pandemic took hold.”

The CODOP predictor is freely accessible at https://gomezvarelalab.em.mpg.de/codop/.

Original Publication:

Klén et. al.: "Development and evaluation of a machine learning-based in-hospital COvid-19 disease outcome predictor (CODOP): a multicontinental retrospective study"

DOI: 10.1101/2021.09.20.21263794

© Pexels/Andrea Piacquadio