In late January, scientists at DeepMind, Google’s London-based AI unit, gathered to discuss whether there was anything they could do to help fight the brewing coronavirus pandemic. At the time, the spread of Covid-19 was still largely confined to the city of Wuhan, but as case numbers continued to grow exponentially, machine learning experts from London to San Francisco were gearing up to try and harness the power of AI to fight the Sars-CoV-2 virus.
“Our first reaction was to think how we might be able to help,” says Demis Hassabis, CEO and co-founder of DeepMind. “Front of mind was our system, AlphaFold, which we had shown could predict the 3D structure of proteins with unprecedented accuracy compared to other computational methods.”
At the start of March, DeepMind released predictions generated by AlphaFold for the structures of various proteins associated with SARS-CoV-2, to try and accelerate the process of understanding how the virus functions. “Understanding the role of these proteins is crucial to developing treatments for the disease,” says Hassabis. "The structure of one of these proteins has since been experimentally determined, and was found to be in good agreement with our predictions, offering a glimpse of how tools like AlphaFold could better prepare us for a future pandemic.”
On Monday, DeepMind announced that the same AI had made a gigantic leap forward. AlphaFold had cracked a 50 year old scientific riddle, the ‘protein folding problem’, determining a protein’s 3D shape from its amino-acid sequence. This will pave the way for faster development of treatments and drug discoveries for nearly all diseases, including cancer, dementia and even infectious diseases such as Covid-19.
During the Covid-19 pandemic, predicting the protein structures of Sars-CoV-2 is just one way in which AI can be used to tackle the growing emergency. Over the past eight months, AI platforms have taken on an unprecedented role in healthcare, as overwhelming patient loads and staff shortages saw medical facilities look to technology to pick up the slack.
From algorithms attempting to optimise ventilator supplies by predicting which hospitals would be most in need of them at which times, to intelligent triage tools collecting information on patients’ symptoms, medical histories and making recommendations on who should receive emergency room treatment first, machines have never had more influence over patients’ lives.
“Covid-19 has been a huge accelerator for the digital healthcare space with mainstream adoption of different technologies which many assumed would take a decade, happening in a matter of months,” says Yonatan Amir, CEO of Israeli health tech company Diagnostics Robotics, which supplies its AI triage tool to healthcare institutions in the US, Israel and India, and has recently signed large scale contracts with organisations such as the Mayo Clinic.
Over the course of five weeks in March and April, Diagnostics Robotics’ tools triaged 2.5 million patients. Overall, Amir says, the demand for its technology from the healthcare industry is 7.5 times larger than before the pandemic. Such intelligent triage tools have made struggling hospitals both more effective and efficient. One example is the Royal Bolton Hospital in the UK, which used a tool from Mumbai based tech company Qure.ai to speed up the waiting time for the results of chest x-ray diagnoses.
“It gives prompt reports to clinicians before a formal report from the radiologist is available,” says Shaista Meraj, consultant radiologist at the Bolton NHS Foundation Trust. “Clinicians have 24 hour access to chest x-ray reports, which helps them make quick management plans.”
But there remains a need for caution when rolling out this technology. While AI systems can handle an intense workload and never miss crucial details because they’re distracted or tired, they can and do still make mistakes with their predictions because they miss certain subtleties when interacting with a patient or evalsuating their complete set of symptoms.
“When a physician looks a patient in the eyes during a routine examination, he or she can tell if the patient understands or not. AI does not yet have these important nuances,” Amir says. “One of the things we’ve realised is that the accuracy of AI prediction often has to do with how the patient question is framed.”
Amir explains that being able to rephrase a question in different ways so that the patient understands is an ability which current AI systems do not have, but are likely to develop in the coming years. “We did a test which found that when patients were asked if their headache was sudden or if it developed over time, most did not know how to answer the question,” he says. “But when we changed the question to ‘Was it sudden, like a lightning strike to the head?’, the answers completely changed. As AI learns more languages and understands the subtlety of nuance it will grow smarter and become more exact with its triage.”
The question of how much to trust the predictions made by an algorithm has also played out in the realm of repurposing existing drugs to target Covid-19. Back in January, scientists at London tech startup Benevolent AI, began using the company’s ‘knowledge graph’ – a large database of medical information consisting of connections extracted from scientific literature by machine learning – to try and identify existing medicines which could be accelerated into clinical trials.
On February 4, the company published its analysis in the medical journal The Lancet. Most notably it suggested that baricitinib, a small molecule approved to treat rheumatoid arthritis, could be effective against Covid-19. Nine months on, the US Food and Drug Administration (FDA) has granted emergency approval for baricitinib to be used to treat hospitalised Covid-19 patients. A Phase III trial showed that the odds of a patient’s condition worsening was significantly lower when the drug was used.
Olly Oechsle, a lead software engineer at Benevolent AI, says the fast approval is “an important milestone” and adds that it has “progressed at an unprecedented pace, moving from computer to bench to bedside in nine months”.
However, while baricitinib ranks as an AI success story, it is also an outlier. Evelyne Bischof, a clinician and researcher at the Shanghai University of Medicine and Health bet365体育赛事s, says that while there are now at least 81 studies using machine learning algorithms to recommend drugs that could be repurposed for Covid-19, no others have received clinical approval. “On the scientific side, AI has certainly been successful in drug development and repurposing for Covid-19,” she says.
“On the clinical side, we still see few tangible, applicable examples.” This isn’t necessarily AI’s fault. Disappointing results from antivirals such as umifenovir and remdesevir – which were touted early in the pandemic as potential treatments for Covid-19 – have somewhat dampened clinicians’ enthusiasm for drug repurposing. In addition, algorithms still require the lengthy process of randomised control trials to verify their predictions, meaning that it can take many months at minimum for a suggested drug to reach the wider patient population.
Scientists faced similar hurdles trying to design completely new drugs to target the Sars-CoV-2 viral proteins. While DeepMind’s AlphaFold platform correctly predicted the structures of a number of the virus’ proteins, it is still in development and so its predictions required months of experimental verification in the lab before any scientists felt they could actually act on the predictions.
At the same time, the sheer scale of the pandemic has seen other AI solutions become mainstream, simply out of necessity. Billions of doses of Covid-19 vaccines are set to be administered around the globe in the next couple of years, after being rushed through clinical trials at breakneck speed. Medical regulators are worried about being able to track reports of safety concerns.
The UK Medicines and Healthcare Products Regulatory Agency (MHRA) has predicted that there could be 50,000 to 100,000 reports of suspicious side effects for every 100 million doses injected. It has awarded a contract to New York based AI company Genpact, to design a platform which uses natural language processing to extract and analyse reports of adverse events which are submitted to the MHRA website from doctors.
AI used on big datasets could also play a role in identifying retrospective clues on why the pandemic affected some people disproportionately more than others, spotting patterns in data which would be beyond human capabilities. Burlington, US based life sciences company LabCorp is using similar technology to mine an anonymised patient data registry consisting of thousands of handwritten doctor’s notes on Covid-19 patients. The aim is to find trends which could provide answers to questions, such as why Covid-19 hit vulnerable socioeconomic groups particularly hard.
If this proves successful in the coming months, trust will grow in such AI tools and they will inevitably play an increasingly important role in medical care. And when the next pandemic arrives, it could mean that we are able to respond much faster and more effectively. “As transformative a technology as AI will ultimately be, this pandemic arrived slightly too early for AI’s current capabilities,” Hassabis says. “If we were unfortunate enough to find ourselves in this situation again, I firmly believe that AI could play a crucial role.”
Demis Hassabis, CEO and co-founder of DeepMind, was a speaker at this year's WIRED Live 2020. The conference on November 24 gathered disruptive minds across technology, design, art and politics to investigate how innovation, technological advances and world events are changing the way we live.
This article was originally published by WIRED UK
