Icon: enlarge
Not a simple formula:
Researcher in the Artificial Research Building at the University of Tübingen
Photo: Sebastian Gollnow / dpa
Artificial intelligence may have already saved tens of thousands of lives in the corona pandemic.
In
any case,
Peyman Gifani
, founder and head of the British company AI Vivo, claims for his algorithms to have given the decisive tip for the most useful drug to date: In mid-April 2020, AI Vivo announced that the steroid dexamethasone was one of five existing drugs Have a chance to help against Covid-19 - the result of an automated search of the Amazon Web Services computers across the medical databases based on a Corona model from AI Vivo.
The larger US-based AI company Nference came to the same conclusion shortly afterwards, which was later confirmed by the Oxford University recovery study.
In fact, before the advent of the newly developed antibody cocktails and vaccines, dexamethasone was the only drug that could reduce mortality.
Since it is abundantly available worldwide and cheap, as it is patent-free, the clinics have been able to make ample use of the knowledge.
According to Nference co-founder
Venky Soundararajan
, the power of human brains is simply not enough to handle the amount of data needed to respond quickly and effectively.
"It's easy to humiliate," he told the Washington Post.
"What we know is just one atom in the universe out there."
If the virus had hit the world 20 years ago - without artificial intelligence - it would probably be over, says
Jason Moore
from the Center for Biomedical Informatics at the University of Pennsylvania, which brings together an international consortium for corona data.
"I think we have a fair chance today thanks to artificial intelligence and machine learning."
An example is provided by the Sandman.MD tool from the Berlin company App @ work, which is now being
optimized
in an EU research project in several university hospitals under the direction of the Frankfurt physician
Kai Zacharowski
for the better treatment of corona patients: All patient data is stored on iPads at the hospital bed recorded and analyzed with artificial intelligence.
By comparing numerous courses, the model is intended to predict how the individual prognosis will develop and which treatment will work best.
No AI model has really been tried and tested
However, one should not expect too much from the technology, warns the medical journal "The Lancet" in a current editorial.
The crisis is forcing the world to use AI models on a large scale "before their usefulness has been proven" - and that could easily backfire.
For a study in the "BMJ", 107 different AI models, which for example predict corona risk groups, diagnoses or predictions, were evaluated.
Many of these would be touted with solid results.
"But all of these models carry a high risk of systematic bias", so the conclusion.
Often the control groups are not selected representatively, the critical patients leave the study before completion, and the models are overly trimmed to provide suitable results.
The claimed performance values are therefore "probably optimistic and misleading".
Sometimes this is also admitted by the AI makers.
Oxford researcher
Andrew Soltan,
for example, and his team have developed a screening called Curial AI, which uses patient data to determine the likelihood of whether a patient arriving at the hospital actually has Covid-19.
The goal: not to even include false positive cases in corona wards and concentrate resources on real corona patients.
According to a study of more than 100,000 British cases, Curial AI was just as accurate, but significantly faster than the conventional PCR corona tests.
But that cannot be generalized to all countries and ethnic groups, warns Soltan.
"We also don't know whether the AI model can be transferred to patient groups in other countries where patients with other health problems visit hospitals."
Without medical application, but still with a corona background, there was probably the first anti-AI uprising in the summer of 2020: British school leavers successfully protested against the awarding of final grades using an AI model that was supposed to replace the missing exams with predicted performance - fair and incorruptible, as the London Ministry of Education assumed.
But as a result, especially children from working-class and migrant families did not get the longed-for university places.
"Fuck the algorithm" became the rallying cry of the protests, a synonym for artificial intelligence without social skills.
Ultimately, the prognosis of the models depends on which parameters the programmers enter and how they weight them.
Even heavyweights like Google or IBM disappoint
Transferred to health: There are AI programs that, for example, are supposed to recognize corona infections by the sound of the voice.
But can you also differentiate between pneumonia caused by Covid and pneumonia caused by the influenza virus?
So far, it has hardly been possible to verify this, as the flu epidemic seems to be canceled this year.
It is precisely because of such problems that even AI heavyweights like Google's sister company X or the IT giant IBM with its super brain Watson have found it difficult for years to deliver the promised solution to health problems using artificial intelligence.
Google wanted to detect flu outbreaks as early as 2008, and Watson would automate cancer patients' treatment plans years later - promised too much.
Meanwhile, the industry has largely shifted to using AI to fish signals from huge amounts of data.
Whether they mean something and what follows from it, then still has to be decided by human intelligence.
ak
