Does AI hold the key to stopping the next pandemic?

8 minute read

Can Australian researchers using AI help stop pandemics before they’ve begun?

Artificial intelligence is a powerful tool for tracking infectious diseases and epidemics, and Australian researchers are on the cutting edge of early detection and tracking. 

But AI has “critical limitations” in public health surveillance, US researchers say in new research. 

“Artificial intelligence (AI) tools can now identify intricate, previously invisible data structures, providing innovative solutions to old problems,” the researchers wrote in the NEJM

“Together, these advances are propelling infectious-disease surveillance forward.”  

But AI can’t replace collaboration between jurisdictions and nations, they said.  

The speed and devastation of the covid pandemic highlighted the importance of “an equally nimble, expeditious, and clever armamentarium of public health tools to counter those effects”, they said. 

“Throughout this crisis, we have witnessed a multitude of AI solutions deployed to play this role – some much more successful than others.  

“As new pathogens emerge or old challenges return to command our attention, the incorporation of the lessons learned into our public health playbook is a priority.” 

AI and machine learning give us early-warning systems, hotspot detection, epidemiological tracking and forecasting, and research allocation. 

But AI has its limitations, the authors said. During the covid pandemic, the success of our current methods was “highly variable”, and future success depends on the limitations of our algorithms, learning from past mistakes and being able to generalise successes.  

There’s also the issue of privacy, which needs to be balanced with the need for high-quality data and “the invasiveness of AI tools”, the researchers said.  

Professor of AI and ubiquitous computing, Professor Flora Salim, said ideally a pandemic would be tracked early to prevent further spread.  

But in those early days and weeks of an outbreak, data is limited – but that’s the time to act to prevent further spread, she said.  

“We don’t want it to escalate further,” said Professor Salim, Cisco Chair of Digital Transport at the University of NSW, who is leading an Australian research team in partnership with a US team on the CSIRO-NSF project into predicting the spread of infectious disease and to understand how bias in AI models spreads to modelling pipelines and leads to biased solutions.  

“But the problem is, an outbreak only escalates further to an epidemic or pandemic when you have more cases that you can no longer control and the infection rate is higher. And of course, we don’t want to get to that stage. We want to tackle it early, but that means we won’t have enough samples to build strong predictive models.” 

Professor Salim said when data was limited, the key was to combine techniques such as self-supervised learning, which leverages pretraining and augmentations of unlabelled data – an approach that’s now prevalent in foundation and large-language models – along with agent-based modelling, which simulates how humans move and interact in a facility or city.  

“If that happens without any restriction what would the spread of disease look like?” 

Professor Salim said well-established mathematical models to track infectious diseases use lot of assumptions about the number of people within a facility, how close together they are in order to model the disease spread, and the reproduction rate, derived from the percentage of infections and transmissibility in that population.  

But combining different data sources and models is a powerful tool, she said. 

“When we fuse them together, data from different sources, including from human mobility, phylogenetics, and web data including social media, is very powerful in improving prediction of cases. 

“Although people keep saying that covid is over, I think this is the time where we have to learn the lessons from covid because there could be very likely new outbreaks and we should be well-prepared in order to avoid another pandemic. 

“How do we make sure that we can detect that risk early on before it becomes a pandemic? That is really the question, especially when we don’t have much data available. How do we augment the decision-making with sparse, limited, noisy data from different countries across the globe, accounting for the biases in our underlying data and fusion it in a way that is responsible?”  

Professor Salim is a professor of AI and ubiquitous computing – in other words, computing that’s a part of everyday life – and researches the quantification of human behaviour using sensor data, and is developing machine-learning analytics techniques to understand human behaviour, underlying activities and intent and patterns from individuals to groups to cities.  

Professor Salim’s project uses spatial temporal data – showing how people move – to reveal how the movement of people relates to early outbreaks among individuals, groups and cities. 

The researchers monitor behaviour patterns and bio signals to show early signs of illness and combine that data with internet search terms around symptoms.  

Professor Salim is also working on an app that uses audio signals to detect “fine-grain behaviour” such as sleep, physical movement and cough sounds associated with respiratory diseases such as covid, COPD, tuberculosis or sleep apnoea.  

Down the track, disease could be detected early through biomarkers, with data gathered by smart watches or smart phone. But that raises privacy issues, Professor Salim said.  

“How do we do this in health or pandemic surveillance, in a very privacy-preserving manner but also very useful manner? This is still a very open-ended question. A key approach we’re taking as a step forward is to enable AI model training on the edge, therefore your data will not need to leave your device.”  

Infectious disease epidemiologist and senior lecturer at the University of Queensland, Dr Amalie Dyda, said there had been an “explosion” of AI and public health research since the covid pandemic hit. 

“With any large-scale event, technology tends to move faster. So there’s really been an explosion of AI and public health in the last two to three years.” 

Dr Dyda said one of the biggest impacts of AI on public health has been its ability to detect early warning signals, and to identify clusters or patterns in disease outbreak at a population level.  

As with any statistical model, the analysis is only as good as the data that’s being used, said Dr Dyda, who is currently researching how technology and machine learning methods can help the public health response to infectious diseases. 

“One of the big challenges for public health and AI is the data sources we’re using because they all have limitations.” 

For example, an analysis using social media data needs to take into account who uses social media and whether that group is representative of the whole population, she said.  

“One of the other big data sets that is quite often used is electronic medical records – and again, who is actually represented by that data? The model is only as good as the data.”  

But the adaptability of AI systems is one of its major strengths, Dr Dyda said. “As data inputs improve over time that the abilities of these systems will improve.” 

Covid also helped push AI development from theory into practice, Dr Dyda said. 

“A lot of AI work prior to covid wasn’t being implemented, it was theoretical, so it bridged those gaps between the people doing these analyses and building these models, and the people working in public health on the ground. 

“While the improvements in implementation are progress, there is still much more work to be done to better incorporate these methods into routine practice” 

Dr Dyda said upskilling the workforce and bridging those gaps between public health and computer science will be crucial in the next decade. 

“It’s really important for clinicians and people in public health to have a voice at this table.”  

Dr Dyda said it was important for those who will be using the systems to have input into how they are developed.  

“The key to successful implementation is ensuring it meets the needs of the users. 

“We need to be considering equity and representation of the data, and issues of safety and unintended consequences. 

“Having people from a public health perspective being able to have a seat at the table and contribute to these conversations about how these things are developed is really important to upskilling our workforce. It’s something we need to focus on in the next 10 years.” 

NEJM 2023, online 27 April 

End of content

No more pages to load

Log In Register ×