Until the coronavirus pandemic took hold, it seemed that at least every other headline about healthcare innovations was related to machine learning or other forms of artificial intelligence (AI) – systems that can mimic human processes such as the capacity to learn and adapt on the basis of new information increasingly used in technology.
We have heard a great deal about how AI might improve healthcare, but what use is it in the face of the massively disruptive effects of a serious infectious disease outbreak?
Monitoring the virus
In fact, AI is supporting a whole range of efforts to manage COVID-19. It has been reported that an AI-driven surveillance system identified the outbreak before even the World Health Organisation knew of it, based on data from a mass of sources including reports from health, public health and livestock bodies, and even allowed prediction of likely spread to other countries. AI continues to be used to track real-time spread around the world.
Identifying interventions
Certainly, the beauty of AI is the capacity to make sense of vast datasets that exceed human analytic ability; combining AI with genomic sequencing is, unsurprisingly, a powerful biomedical tool. One application is the analysis of viral genome sequences and combine these with computer modelling to predict potential responses of the virus to different drugs. AI is used to model the interaction of the virus particle with a potential drug in silico to identify candidates that may be effective at disabling the virus.
US researchers have used AI to identify a potential inhibitor of viral-induced lung disease, baricitinib, now in clinical trials. Many other companies and research organisations are using similar approaches to screen known drugs and other chemical compounds that could prove to be useful treatments, sometimes using AI to search the vast amounts of information available in the published scientific literature, as well as libraries of compounds.
Modelling can be used to predict the physical and biological changes that could arise from changes in the viral genome sequence – increasing or decreasing the ability to infect humans, to cause disease, or to interact with potential treatments. This could provide early warnings of trouble, or suggest changes to any new treatment approaches that might be found. AI-driven modelling is also being used in the vital search for an effective vaccine; systems such as DeepMind’s Alphafold are being used to model the precise physical structure of the virus, to help the design of vaccines that can perfectly mimic key parts and trigger protective immune responses.
Planning and personalising care
AI can also be put to good use with many other forms of data. In the UK, researchers at the University of Cambridge are using machine learning to analyse COVID-19 patient information from Public Health England in order to predict the risk of patients developing more severe disease – and needing specific resources such as intensive care beds and ventilators. The mathematicians are using information on patients being admitted to hospital, to intensive care units, needing respiratory support, and whether and how fast they recover, to develop predictive algorithms that could help hospitals plan ahead for likely resource demands.
It may even be that this sort of algorithm could be developed to aid clinical decision-making for COVID-19 patients – for example, identifying patients at risk of rapid deterioration to receive extra monitoring or earlier respiratory support. Another personalised medicine benefit might lie in suggesting those patients for whom emerging treatments are likely to be more or less effective.
Supporting health services
On the frontline of health services AI is helping with the delivery of healthcare, just as they are aiding other services. The sudden shift to remote digital health services for most patients, to reduce the risks of COVID-19 transmission, has seen AI-driven health apps gain further prominence. These may be directed to automating basic (but essential) processes such as taking patient information, or powering remote consultations and triage such as the Doctorlink symptom assessment platform now being used for NHS video consultations. The UK government has also launched a WhatsApp AI-powered ‘conversational AI’ chatbot to provide coronavirus information to the public.
One important factor in managing the current outbreak is knowing who is actually infected. Current diagnostic tests use a technique called RT-PCR, but Chinese researchers have reported using AI to analyse CT chest scans in a bid to see whether or not COVID-19 lung disease could be distinguished from other, similar chest infections. Results appear promising , though concerns have been raised about potential selection bias.
The perennial data challenge
Success against COVID-19 from AI relies on two factors: the AI tools themselves (in terms of quality and performance), and crucially, the quantity and quality of the data used to develop and train those tools.
As Prof Ara Darzi, director of the Institute of Global Health Innovation, at Imperial College, told the BBC: “AI remains one of our strongest paths to achieve a perceptible solution but there is a fundamental need for high quality, large and clean data sets” – or as the old adage goes, rubbish in, rubbish out; issues such as inadvertent bias introduced by the data used to train AI systems also require careful scrutiny. This is why regulating AI systems as medical devices is no simple matter.
There is little question that AI is supporting the fight against COVID-19 on many fronts; it may yet prove to have played a vital role in success, by aiding the identification of an effective new treatment or vaccine. However, AI is not the whole answer; the true benefits lie in combining AI tools with human insights and experience to deliver optimal public health and medical intelligence and interventions.