They use artificial intelligence to predict the survival of Covid patients in intensive care
Health systems around the world are struggling to care for the large number of serious patients with Covid who need special medical attention, especially if they are identified as high risk.
In the intensive care units scales or forecasting tools are often used, such as SOFA The APACHE II, to predict the evolution of patients based on various parameters, but its reliability is limited in the case of covid-19.
Now European scientists have shown that blood samples of a seriously ill patient due to this disease, specifically the proteins in their blood plasma, can be analyzed with a machine learning model to predict weeks in advance whether or not the person will survive. The results are published in the open access journal PLOS Digital Health.
“Our study shows that a combination of proteomic markers, combined in a risk prediction model based on artificial intelligence, can predict quite well the probability that an individual patient will die or survive covid, ”says co-author Florian Kurth Charité University Hospital in Berlin (Germany).
“Furthermore,” he adds, “the proteomic risk prediction was much better than the prognosis derived from risk assessment scores commonly used in clinical care.”
To conduct the research, the authors began by studying the levels of 321 proteins in blood samples taken at 349 times or points in time in 50 patients with covid-19 in critical condition who were being treated, with mechanical ventilation, in two health centers in Germany and Austria.
Then used the machine learning to find associations between measured proteins and survival of sick people. Of the cohort or trial group analyzed, 15 patients died and the mean time from admission to death was 28 days. For those who survived, the median hospital stay was 63 days.
Using blood tests, the researchers identified 14 proteins (such as alpha-2 macroglobulin, APOC3, GPLD1, various serpins…), whose measurements changed over time and moved in opposite directions on the graphs depending on whether or not the patients survived in intensive care.
“Interestingly, the Plasma levels of all these proteins had previously been altered by the disease, depending on the severity, which makes us confident in our findings,” says Kurth, who explains: “The proteins with the greatest relevance in the prediction model belong to the coagulation system and to the so-called cascade or plugin system (a component of the immune response). Both are known to be especially important for the pathophysiology and severity of COVID-19.”
The team then developed their machine learning model to predict survival from a single temporal measurement of the relevant proteins, and tested it in Austria with 24 patients in critical condition because of the coronavirus.
For this group, the model demonstrated a high predictive power, correctly predicting 18 of the 19 patients who survived and five of the five people who died.
The researchers conclude that blood protein tests, once validated in larger cohorts, can be useful both to identify patients with a higher risk of mortality and to better understand the disease and check whether a certain treatment changes the prediction when applied to individual cases.
“Now we want to see if we can transfer this methodology from research facilities to an everyday environment, to a standard clinical measurement laboratory, in addition to evaluating the method in larger groups of patients, and possibly also for other diseases”, says Kurth.