Female doctor checking on Covid-19 infected patient while connected to a ventilator
Female caucasian doctor checking on Covid-19 infected patient while connected to a ventilator at a hospital room

A group based at Charité Medical University in Berlin has developed a machine learning based tool that can predict who will be most ill as a result of COVID-19 based on the results of a blood test.

The tool can predict patient survival based on levels of a range of blood plasma proteins on admission to hospital, something the researchers hope could save lives by highlighting who needs the most care at an early stage.

Methods to assess the risk for dying of patients admitted to hospital intensive care units are already in existence, such as APACHE (acute physiology and chronic health evaluation) II and Sequential Organ Failure Assessment (SOFA), but these measures have not proved very accurate for seriously ill COVID-19 patients.

To try and improve these predictions and better allocate care to sicker patients, Florian Kurth and Markus Ralser from Charité, along with research colleagues, created a machine learning model from COVID-19 patient blood protein data, based on parenclitic networks comparing one dataset with another.

The researchers first analyzed blood proteins from 50 severely ill patients with COVID-19 who were treated between March and September 2020 at Charité University Hospital in Berlin, 15 of whom died.

“Studying the plasma proteomes, we found 78 proteins for which the concentration changed significantly during the patients’ disease course. Out of these proteins, 14 were found to change differently over time for survivors and non-survivors,” write the researchers in a paper describing the work published in PLOS Digital Health.

Patients who died had higher levels of a group of different inflammatory proteins: SAA1, SAA2, CRP, ITIH3, LRG1, SERPINA1, SERPINA10 and LBP and lower levels of the anti-inflammatory proteins SERPINA4 and A2M. The coagulation related proteins thrombin (F2) and plasma kallikrein (KLKB1) were also lower in patients who did not survive.

“The majority of proteins with high relevance in the model are components of the coagulation system and complement cascade, highlighting their critical role in progression and outcome of most severe COVID-19,” explain the authors.

After creating a machine learning network using this data, the team tested its accuracy at predicting survival in an independent cohort of 24 critical COVID-19 patients from Austria. The model was highly accurate and correctly predicted 18 or the 19 patients who survived and 5 out of the 5 who died.

“We show that the proteome accurately predicts survival in critically ill patients with COVID-19, from samples that were collected 39 days in median before the outcome,” write the researchers.

While they acknowledge the study was small is size and the findings need to be confirmed, they say “the findings warrant further prospective assessment of proteomic predictors and the described models in larger cohorts.”

Also of Interest