A simple blood test has been developed that predicts the patients most likely to die from Covid.
It identifies 14 proteins in the blood associated with survival, scientists said.
The test involves a single sample from a critically ill sufferer that can then be analysed by AI (artificial intelligence).
It uses plasma levels to work out mortality risk with what researchers say is almost 100 percent accuracy - weeks before the outcome.
The breakthrough could overcome one of the big challenges in the pandemic.
The virus is more dangerous in the elderly or those with pre-existing health problems.
But that doesn't preclude a 96-year-old from presenting with no symptoms - or a healthy 21-year-old losing their life.
Co author Professor Florian Kurth, of Charite University Hospital, Berlin, said: "The clinical picture of Covid-19 is exceptionally diverse - ranging from asymptomatic to very serious disease and death.
"For physicians, it is difficult to estimate the individual risk for a patient of deterioration or death."
Study examined 'biomarkers' in patients on ventilation
The international team examined biomarkers in a similar cohort of patients who were all severely ill.
They were in intensive care and required ventilation with additional organ replacement therapy.
Healthcare systems around the world are struggling to accommodate severely ill patients - especially if they are deemed high risk.
Current assessments in intensive care medicine show only limited reliability in identifying the most vulnerable.
The researchers analysed levels of 321 proteins in blood samples taken at 349 time points from 50 individuals being treated in Germany and Austria.
A machine learning method was used to find links between the measured proteins and patient survival.
Prof Kurth said: "We found 14 proteins which over time changed in opposite directions for patients who survive compared to patients who do not survive in intensive care.
"Interestingly, the plasma levels of all the proteins had been found to be altered by Covid-19 before, depending on the severity of the disease.
"This makes us particularly confident in our findings."
The average time from admission to death for the 15 patients who died was 28 days. Survivors remained in hospital for around two months, on average.
The computer neural network forecast survival based on a single time-point measurement of relevant proteins.
They tested the model on an independent group of 24 critically ill patients.
It correctly predicted the 18 of 19 who survived - and all five and 5 out of who died (AUROC = 1.0, P = 0.000047).
The method described in PLOS Digital Health may also be useful in testing whether a treatment changes the projected trajectory of an individual patient.
Prof Kurth said it may even have potential applications for other diseases.
For more stories from where you live, visit InYourArea.
