Why didn’t AI stop Covid-19

In late January 2020, a week before Covid-19 was named, hospitals in Wuhan, China, began testing disease screening using AI. The method selected is a CT scan (CT) of the lungs. Thousands of CT scans will be imported into computers, using algorithms to determine if pneumonia is caused by Covid-19 or common illnesses like seasonal flu.

When nCoV began spreading in the United States in February, this idea seemed very promising, allowing rapid identification of people infected with the virus, especially in the absence of test suites and many of the available products were defective. Still, US health officials are uncertain about that.

The Food and Drug Administration (FDA) has approved many diagnostic algorithms for conditions ranging from bone fractures, eye pain to breast cancer, but they often take months or years to complete. These algorithms must be implemented in many hospitals with different patient groups and undergo non-stop error checks.

Is there enough data about nCoV to distinguish the symptoms of pneumonia due to different diseases? How to identify cases of mild symptoms, when lesions are more difficult to distinguish? The pandemic is still going on, but treatments will have to wait for the full answer.

• Boundaries between AI and human rights 12

In late March, the United Nations and the World Health Organization (WHO) published a report on evaluation of CT scans, along with a series of AI applications for the fight against Covid-19. “Only a few projects are able to put into operation,” the report said.

These limits existed long before Covid-19 broke out, but the pandemic exacerbated the problem. The reliability of AI depends on the ability of humans to collect and analyze data. Covid-19 has become an example of why this is so difficult during a pandemic.

The daily actions are determined by a series of uncertain projections of the group of infected and dying people, as well as the number of deaths if community isolation measures fail. Even the advice on facemasks and which patients are prioritized for breathing machines is constantly changing, making it difficult for data collection efforts and pandemic forecast.

“AI is always behind people in Covid-19, while many people think that it has the ability to predict beyond what we expect,” journalist Gregory Barber of Wired commented.

The process of drug development is another example. One of the most noteworthy tests is that of Google DeepMind. The company’s AlphaFord system is leading the field in building protein models and predicting the basic structure of viruses. The normal lab analysis process can take months, while DeepMind only needs a few days. The team said it was an estimated model developed by a system that was still in beta testing, but it gave the impression that AI had entered the race to develop a vaccine against Covid-19.

Even so, the medical community is not very impressed. “I don’t see the true role of AI at this time,” commented Julia Schaletzky, director of the Center for Emerging and Forgotten Diseases at the University of California-Berkeley, USA.

She said that many protein samples have been identified in the lab without AI and it would be risky to spend time and effort to start from scratch and use the incomplete system. “Technological progress is a good thing, but it often has to be exchanged by abandoning the option to build on the existing and full potential,” Schaletzky said.

Experts say that AI has potential in finding treatment options. AI algorithms can complement data collection techniques to filter a wide range of available information, such as promising research or past effective solutions. One of the drugs found in this way is Baricitinib, which is undergoing clinical trials.

The AI ​​can also analyze how nCoV attacks the body, in which the algorithm will exploit patient data to identify people at high risk of death and who have the ability to recover, thereby developing a treatment regimen. complete.

However, all factors still depend on the input data, which is what humans collect and arrange for AI to analyze. Health care systems in other countries do not easily provide patient information to develop such systems, and a series of privacy regulations and incompatible databases will hinder AI researchers. .

“The current crisis can change that, even motivate us to rethink the way we store and share data. We may continue to study nCoV after the pandemic has ended, in order to prepare better algorithms before the next pandemics. Still, it would not be surprising if AI could not save us from Covid-19, ”Barber commented.