Researchers are looking at how artificial intelligence (AI), including large language models like ChatGPT, could be used to scenario plan for future epidemics.
What if you could interpret large amounts of health data faster to predict how long a patient might stay in the hospital, or input human behaviour into an epidemic model to determine the possible curb of a viral outbreak?
These are some of the ways researchers are testing new artificial intelligence (AI) models to better plan for future viral outbreaks such as “Disease X,” an unknown pathogen that could launch an epidemic akin to COVID-19.
“One of the strengths that we're seeing with AI-based approaches to analysing large datasets is really the ability to identify early signals of potential anomalies in the public's health,” Alain Labrique, director of the Department of Digital Health and Innovation at the World Health Organization (WHO), told Euronews Next.
“I think there are many different ways an advanced computational tool like artificial intelligence can be used to enhance the way we detect new epidemics and pandemics, but also respond to those epidemics and pandemics”.
But he added that addressing bias and giving the models good data, not just from one specific population, is important to strengthen them. It’s a field where research is increasing but in practice, some of these new models could take time to roll out.
Disease severity and hospital capacity planning
Researchers at Yale University in the US recently published a study that addresses one of the many challenges that arose during the COVID-19 pandemic: how to manage overflow in hospitals.
“The number of hospital beds is limited and if you have a pandemic like [COVID-19], you want to prepare. We're looking at the public health point of view. We want to get prepared if something happens,” Vasilis Vasiliou, Chair of the Department of Environmental Health Sciences at Yale School of Public Health, told Euronews Next.
Their epidemic model uses an AI-powered platform to triage patients by predicting how bad the disease will get and how long they could spend in the hospital.
It’s based on clinical and metabolic biomarkers that they found helped to indicate how the disease would progress.
Vasiliou says that in a future viral outbreak, the focus would be on getting early data into an AI-powered algorithm that works to determine how to better organise hospital resources.
“If something comes up very quick, you have the framework, you have a model, you have an algorithm that you feed immediately [with] the first data from the first country where this has happened. Then you can start building up a new model,” he said.
For him, one of the current limitations is a lack of data. “Every AI model, the more data that you feed into it, the fewer limitations you’re going to have,” he said.
Kirill Veselkov, a co-author on the study from Imperial College London, said that with an emerging disease, you need to find new biomarkers that can influence its severity.
“The current state-of-the-art analytical tools will be capable of measuring hundreds of thousands of these biomolecules,” he said.
“So, if you want to analyse them, it's probably going to be impossible by human doctors without the use of sophisticated mathematical algorithms and AI is particularly suitable for that, to identify the pattern or set of biomarkers and associate them with the disease process and disease outcomes,” he added.
But their model will need to be further studied with more populations, taking into account comorbidities and other factors before it could be generalisable for the larger public.
‘Using AI to work out when to lockdown’
For COVID-19, which is a virus that we already have information about, AI can help with hospital scheduling, according to Rachel Dunscombe, a UK AI council member and current CEO of OpenEHR.
“What we've got is a set of data on the ground that will tell us the real-world situation and we need to know whether we need to put interventions in, if we need to lockdown, if we need to increase the capacity of the systems, if we need to, you know, reduce elective activity to make space,” Dunscombe, who is also the former CEO of the NHS Digital Academy, told Euronews Next.
“We can actually use AI [in healthcare planning] to work out when the appropriate time is to lockdown, to put masks in place, you know, to put additional staff in place to reduce the activity we do day-to-day,” she added.
She said in the UK, they feel more equipped to actually use models to work out the impact on the ground of certain scenarios after the COVID-19 pandemic.
“If it's fed the right data and it's supervised in the right way, it will give us the likely outcomes,” she added.
‘Hard to represent human decision making’
Researchers at Virginia Tech in the US are trying to solve a separate epidemic modelling problem using AI; that is how to accurately represent the complexities of human behaviour during a viral outbreak.
“In the traditional modelling, you have to somehow represent human decision making” which is hard to do, Navid Ghaffarzadegan, an associate professor at Virginia Tech, told Euronews Next.
“The reason is that humans are complex. Societies are hard to predict. With better or different ways of representing humans through AI, you now have the ability to see how they react under different scenarios, and you have models that are incorporating human behaviour in them,” he added.
As part of their study, which is currently in pre-print, the researchers modelled an epidemic in a town named Dewberry Hollow with a fictitious virus called Catasat, in order to avoid possible biases when using ChatGPT.
They explored how humans deciding to stay at home or not could influence the epidemic model by providing a scenario and personality characteristics of different pretend “agents”.
They found that these generative AI-empowered humans mimicked “real-world behaviours such as quarantining when sick and self-isolation when cases rise” in the simulation.
The multiple waves of the virus were similar to waves seen during previous pandemics that ended in the virus becoming endemic in society.
The main limitation, they say, is that it is costly and time-consuming to run, though they expect that as AI continues to develop this could improve. Others say that their model still needs validation.
The future of AI and pandemics
In a separate paper published last year, Ghaffarzadegan highlighted the difficulties of forecasting the trajectory of an epidemic with both traditional and AI models. He found that AI models did not necessarily perform better, but said that it’s in part due to changes in human behaviour.
Some say there’s still little research assessing the performance of AI during the COVID-19 pandemic.
A review article published in the Frontiers in Medicine journal in 2021 analysed 78 studies regarding AI use during the pandemic.
Its uses included AI-assisted diagnosis for COVID-19, epidemic prediction as well as drug development, such as rapidly identifying drugs or products that could neutralise the disease.
They concluded it was a potential tool during epidemics but that continuous research on it was needed.
Veselkov said the AI-powered triage study was in the research and development stage but that it would still take time before these AI models could be used to plan for future epidemics, such as an unknown pathogen WHO calls Disease X.
“We need to really develop the tools, but also think a lot, especially when it comes to healthcare applications, pandemic applications, population level applications, we need to think about the safety and the robustness of the solution as well as the limitations of the solution,” he said.