How the Centre’s network of leading scientists and clinicians is deploying pioneering AI and machine learning on multiple fronts to help fight the coronavirus pandemic.
The coronavirus pandemic has galvanised scientists, clinicians and technologists in a global attempt to save lives, mitigate further economic devastation and bring this pandemic under control at the earliest opportunity. The Cambridge Centre for AI in Medicine (CCAIM) is at the forefront of these efforts, bringing pioneering AI and machine learning (AI+ML) to bear on the challenges raised by the pandemic.
COVID-19 is creating battles on innumerable fronts and all at scales. For people infected with COVID-19, the fundamental fight is cellular. For hospitals, the struggle to protect their patients and staff continues while administrators marshal precious resources, such as ICU beds and ventilators. For scientists racing to find effective treatments or vaccines, the fight is often in silico, as computer-modelled drugs are put through their paces.
CCAIM is mobilising its internationally recognised expertise on a number of fronts: from using advanced modelling techniques to help uncover new drug targets against the novel SARS-CoV-2 virus, which causes COVID-19, to partnering with Public Health England and the NHS to develop and trial multiple AI-based systems to forecast ICU-capacity as the epidemic evolves.
Meanwhile, the Human Cell Atlas international consortium, founded and principally led by CCAIM team member Dr Sarah Teichmann of the Wellcome Sanger Institute, has moved quickly to analyse existing and new data to develop insights into the biology of the coronavirus, and to release these data and analyses openly, and with the urgency the pandemic demands.
The early months of the pandemic in the UK saw scientist working fast to calculate the impact of the anticipated surge of COVID-19 patients into hospitals, and the anticipated pressure suddenly placed on limited intensive care beds and ventilators. Even in “normal” times, the occupancy of ICU beds in the UK is typically greater than 80 per cent, and occasionally maxed out, so the looming situation was clearly precarious at best, particularly in the light of how, in March, coronavirus was overwhelming many of northern Italy’s hospitals.
In the vanguard of researchers applying machine learning to predict demand on England’s ICU capacity was CCAIM’s Dr Ari Ercole from the Division of Anaesthesia at the University of Cambridge. Working with CCAIM colleague, Professor Pietro Lio, a computational biologist at department of Computer Science at the University of Cambridge, and others, Dr Ercole demonstrated how publicly available data can be rapidly combined to dynamically model short-term ICU demand as the COVID-19 pandemic unfolded. The researchers’ study, released rapidly and with open data on 23 March – and picked up by national media – suggested that UK’s ICU capacity was at short-term risk of being overwhelmed.
“For clinicians, watching this new disease spread rapidly in front of us has been a huge challenge and has required reconfiguring all aspects of healthcare. Data science and modelling is crucial in allowing us to deal quickly with the uncertainty of an emerging disease,” says Dr Ercole. “At an operational level, early forecasting of ICU demand has allowed the country to expand capacity in an appropriate way and this has prevented resources being overwhelmed as has been seen elsewhere in the world. Only through timely data is it possible to stay ahead of this pandemic.”
Only through timely data is it possible to stay ahead of this pandemic.
On 27 March, a few days after that study was published, the CCAIM team, led by Centre Director Professor Mihaela van der Schaar, issued a rallying cry to governments and healthcare authorities to embrace proven AI+ML techniques, in combination with the wealth of existing and rapidly emerging data, to coordinate a response to the COVID-19. The agenda-setting paper, “How artificial intelligence and machine learning can help healthcare systems respond to COVID-19”, laid out key challenges in five areas – the management of limited healthcare resources; the development of personalised patient management; informing policies; accounting for uncertainty and expediting clinical trials – and, crucially, specified how AI could be quickly applied to address these issues.
Several weeks later, on 20 April, Professor van der Schaar, NHS Digital and Public Health England announced an exciting partnership. Professor van der Schaar’s Cambridge lab had adapted a system they had recently developed, called Cambridge Adjutorium, to make hospital-level projections of upcoming demand for ventilators and ICU beds, and successfully trained the system using depersonalised COVID-19 patient data provided by Public Health England. The system’s predictive accuracy far surpassed existing state-of-the-art techniques. NHS Digital is trialling Cambridge Adjutorium at a number of Acute Trusts in England.
Professor Van Der Schaar is an engineer, and she leads a multidisciplinary team that develops, tests and delivers working new solutions to the really hard problems in medicine.
“Professor Van Der Schaar is an engineer, and she leads a multidisciplinary team that develops, tests and delivers working new solutions to the really hard problems in medicine,” said Dr Jem Rashbass, Executive Director for Master Registries and Data at NHS Digital, when the partnership was announced. “Two weeks ago, the team shared a method with the world that showed it was possible to do capacity planning for COVID-19 patients. We recognised that there was an opportunity to industrialise the methods and deploy this as a service through the national infrastructure managed by NHS Digital and deliver a real data-driven planning tool to hospitals.”
NHS planning support
Another CCAIM team member is exploring similar territory, using time series and system dynamics (SD) models to support regional COVID-19 capacity planning. Stefan Scholtes is Dennis Gillings Professor of Health Management at the University of Cambridge’s Judge Business School. Along with other faculty and PhD students, he has worked since early March with Public Health England and the NHS in the East of England on forecasting hospital and ICU bed capacity requirements for COVID-19 patients.
In the early phase of the epidemic, prior to the first peak, the group tested a variety of time series models. Even with relatively little data, these models performed well when forecasting 7 to 14 days ahead, outperforming less sophisticated models based on the exponential increase in COVID-19 patients. The team’s superior time-series models were updated daily and fed into the regional NHS capacity planning process.
In parallel, Scholtes’s team developed an SD model that integrates a traditional epidemic model with a testing module and a hospital module to capture the dynamic interplay between supply and demand. As the epidemic has evolved, the SD model’s performance has improved markedly, resulting in it outperforming the pure time series models and providing more accurate and reliable longer-term forecasts. A combination of time series and SD models has proven powerfully predictive.
The SD model is now being expanded beyond COVID-19 patients to inform the allocation of NHS capacity between COVID and non-COVID patients during the restoration phase of the pandemic.
Smarter drug targeting
Of course, one way to prevent people infected with COVID-19 from needing intensive care in the first place is to find effective drug treatments for the disease. This is where ongoing work by a team led by CCAIM’s Co-Director, Andres Floto, a Professor of Respiratory Biology at the University of Cambridge, comes in. Bridget Bannerman in his team is exploring a combination of structural and dynamic modelling approaches to predict new drug targets against SARS-CoV-2. This methodology involves the analysis of a metabolic model that integrates the cell metabolism of various cells of humans infected with coronavirus.
The model is providing an in-silico comparison of the biochemical demands of the virus versus the infected host cells. The team is now expanding the model to predict the effect of various treatment regimens for COVID-19, aiming to find the best drug optimisation strategies against the virus.
The more we can learn about the fundamental biology and interactions of SARS-CoV-2, the more powerful this sort of drug discovery work becomes. Which brings us to CCAIM team member Dr Sarah Teichmann, and her colleagues at the Wellcome Sanger Institute and Human Cell Atlas (HCA) Lung Biological Network.
To explore which of the body’s cells could be involved in the transmission of COVID-19, these researchers analysed multiple datasets of single-cell RNA sequencing, from a variety of tissues from healthy people, including cells from the lung, nasal cavity, eye, gut, heart, liver and kidney. The research, published in Nature on 23 April 2020, looked for individual cells that expressed both of two key proteins that the virus attaches to when infecting our cells – ACE2 and TMPRSS2. They discovered that two types of cells in the nose have particularly high levels of these, making them likely points of entry for COVID-19. This sort of science is foundational to developing medical responses to the pandemic.
From fundamental science to capacity planning in intensive care departments, the threat of coronavirus requires that we use every weapon in our scientific arsenal, and continue to develop more powerful ones as we learn more about the virus and the disease. For its part, the Cambridge Centre for AI in Medicine’s expert network will continue to bring pioneering AI+ML to bear on COVID-19.