AI And Coronavirus?

The answer to the question, “Is artificial intelligence being used in X?” these days is often yes. If you searched for information online, listened to streaming music, browsed Netflix and didn’t receive any spam recently, then you used AI. It’s important to understand this, because there’s been a lot of hype around AI being deployed as a weapon against the coronavirus pandemic. It obviously is.

“When people talk about AI they often refer to it as a technology, but for me it’s not a technology; it’s an enabling technology,” says Deshen Moodley of the Centre for Artificial Intelligence Research. “If you have something that’s already digital, AI can enhance it. Practically, at the moment, there are mature applications of AI that have radically changed our lives. The thing is, these are old and established so most people don’t think of them as being AI-enabled – like the search technology behind Google, for example.”

Aiding contact tracing

AI is also behind the route-planning capabilities of mapping apps. Understanding the movement of people is at the heart of contact tracing, as public health officials attempt to identify and monitor people who have been exposed to the virus, in order to understand the source of their infection and limit further spreading of it.

Moodley gives the example of Uber. “At any point in time, the app is tracking both drivers and customers – say a few hundred thousand people across a country. Uber knows when these two entities come into contact because they have phones.

“Now, if you take that and just make those points to people, you understand how contact tracing can work. You can also introduce a proximity element to it too, to tell if people have come close. The accuracy of GPS is only about 1.5-2m, but you can estimate this using algorithms – say two people pass each other on a main thoroughfare or walkway, then you have a dimension and can probably tell with some probabilistic measure if they’ve come into close contact.”

That’s pretty much exactly what Singaporean tech company SQREEM Technologies has done in repurposing its AI-driven digital advertising analysis software into a contract-tracing system, which has been deployed in South Africa. It uses publicly available cellphone metadata to track population movement across a geographic grid. When someone enters a square where the likelihood of coming into contact with COVID-19 is high, the government will target their web activity with ads directing them to health services, and may even retrieve their phone number from the service provider to send them SMSes directly. SQREEM itself doesn’t actually handle the data; this is done by the authorities only.

Taming the literature

Scientists don’t know how to treat or vaccinate against the novel coronavirus just yet, but that doesn’t mean they’re not trying. Nature Index reports that at the beginning of March, there were 2 460 COVID-19 papers published; by 20 May that number sat at 36 883, with around 5 000 new papers being published every week. That’s on top of the tens of thousands of existing papers related to coronaviruses.

While it’s vital that research continues apace, medical authorities, researchers and physicians are faced with attempting to sift through this embarrassment of riches for pertinent information. One initiative with the aim of helping them see the wood for the trees is the COVID-19 Open Research Dataset (CORD-19), created by the Allen Institute for AI in partnership with the White House Office of Science and Technology Policy and a number of leading research groups.

Launched on 16 March, CORD-19 is a collection of more than 120 000 papers on SARS-CoV-2 and other coronaviruses, formatted to be more easily searched by algorithms. Lucy Lu Wang, one of the lead data scientists on the project, says they’ve seen two main types of usage so far. “The dataset itself was produced primarily to assist with text-mining, natural language processing (NLP) and these sorts of AI applications,” says Wang, “but we’ve also seen some use outside of this intended area by medical and clinical practitioners. So it’s being used as another source of scientific literature to conduct things like systematic reviews and study biomedical topics – drug repurposing, interactions with organ systems.”

Wang says that hundreds of live systems have been built on top of CORD-19 to access the data in its documents. “We’ve seen things like language models tuned to these documents, knowledge graphs, entity labelling. But the live systems are potentially the most useful – we know of more than 60 sites where the documents have been indexed so anyone can search through them and ask questions. A lot of these systems are partnering with medical doctors to get a better sense of how effective these search systems are.”

A number of shared tasks and competitions have sprung up around the dataset, particularly with regards to asking questions and getting structured answers from the available data. “The world is focused on coming up with treatments, managing the disease and developing vaccines, so being able to answer questions quickly about what’s been tried, what works and what doesn’t work is one way to speed up these processes,” says Wang. “It can also help establish if we have enough information to abandon a particular idea.”

Towards better risk evaluation

It’s well documented that reactions to the coronavirus are varied; most people experience mild symptoms, but those with serious symptoms can be at serious risk of debilitation or death. Researchers led by NYU Grossman School of Medicine and the Courant Institute of Mathematical Sciences at NYU, in partnership with China’s Wenzhou Central Hospital, are using AI to predict which COVID-19 patients will develop severe respiratory disease.

“Epidemics often create a perplexing problem where there is a surge of patients, a novel disease that doctors and nurses have no clinical experience with, and limited resources,” says corresponding study author Dr Megan Coffee. “Being able to identify early which patients needed the most care would help doctors and nurses best target their care, especially when resources are tight.”

“Most people who are infected do not need to be hospitalised, but about one in five patients who develop symptoms need to be hospitalised,” says Dr Coffee. “Of these patients who require hospitalisation, about one in four requires intensive care – about 3-5 per cent overall. Those who are older, male or have hypertension or obesity may be more at risk of severe disease. But it can be enigmatic who develops this. We wanted to see if the virus left a footprint that could be identified on first presentation in order to know ahead of time who would develop severe disease.”

To this end, the researchers gathered data on symptoms as well as lab values like kidney function, blood-cell counts and X-rays from 53 patients at two hospitals in Wenzhou, China, tracking their disease progression and recovery. From this data, they built a set of algorithms utilising predictive analytics (a form of AI) to match early symptoms with severe later illness.

“In a general sense, predictive analytics learns from experience – in this case historical clinical data – to find patterns, and then uses those patterns to make predictions about the future,” explains study co-author Dr Anasse Bari. “Our ensemble of machine-learning algorithms aimed to determine which mildly ill patients were likely to become severely ill.”

They did so with a 70-80 per cent accuracy rate – close to the standard for use of AI in medicine. The research team continues to collect data from more locations, primarily in New York City, while fine-tuning the algorithms to increase their accuracy. The goal is to provide a decision-support application for doctors.

“We expect to have the tool as an app for doctors that can scan through a new patient’s data quickly and provide a risk assessment for that patient,” says Dr Bari. “We aim at providing explainable AI with evidence from the data that explains the outcome provided. This will not replace a doctor; it is a decision- support tool that will only provide a ‘recommendation’. The doctor will have to make the final decision. The tool will also be open to take feedback, tuning or correction from the user to improve its learning.”

Simulating synthesis
A Polish start-up called Molecule.one is using AI to aid in the creation of new molecules that could be used to treat illnesses, by conducting molecular biology in a high-fidelity simulation. By automating the journey from available compounds to desired compounds, they say they can dramatically speed up the process of drug creation.