What’s the News — AI Boosts Flu Out Break Predictions:
Great news! AI has been used to spot and predict flu infection patterns. This is particularly timely given the ongoing coronavirus pandemic and potential application to other public health crises. Yue Ning, PhD, and colleagues at Stevens Institute of Technology in New Jersey claim that their AI technique produces an 11% boost in accurately predicting outbreaks over that supplied by other contemporary systems, allowing for predictions of flu outbreaks up to 15 weeks ahead of their arrival. As suggested by the Institute, the new technique may help predict local and regional COVID outbreaks.
How is this done? Algorithms Train on Real World Data
The team at the Institute say their success is due to incorporating location data. Past forecasting tools have sought to spot patterns by studying the way infection rates change over time but here the team used a graph neural network to code flu infections as interconnected regional clusters. That allows their algorithm to discover patterns in the way influenza infections flow from one region to another, and also to use patterns from one region to inform its predictions in other locations. The algorithm is trained on real-world state and regional data collected from the U.S. and Japan. In order to assess accuracy, predictions were tested against historical flu data. This approach differs from other models, which generally have focused on past data to forecast flu outbreaks a week or two in advance. Here, the incorporation of location data is touted to allow for better predictions over longer spans of time.
What is a “Graph Neural Network?” — Taking A Google Map Approach.
Graph Neural Networks (GNNs) are deep learning architectures for graph-structured data. That is, they are deep learning models that operate on graphs. At a high-level, the graph data is able to provide rich relation information among elements. “Unlike standard neural networks, graph neural networks retain a state that can represent information from its neighborhood with arbitrary depth.” GNNs act to collectively aggregate information from graph structures, wherein the nodes on the graph are defined by their features and related nodes and the “neighborhood” for each node is calculated. Other known applications GNNs have been used for include traffic predicting in collaboration with real-time location sharing, in terms of maps and the like.
The team at the Institute is now said to be working to improve the influenza-forecasting algorithm by incorporating new data sources. The team, however, is constrained by publicly available information. For example, one challenge the team has faced has been to figure out how to account for public health interventions, including vaccinations, mask utilization, and social distancing practice. “It's complicated, because health policies are enacted in response to outbreak severity, but also shape the course of those outbreaks,” Ning explained. “We need more research to learn about how health policies and pandemics interact.” Having access to this information would better help teach the predictive AI system, but getting others to collect and share the data is a problem. Perhaps one way to get around this is for future public health laws and regulations to take this into account when enacting requirements for public health data collection.
What’s Next in the Use of Public Health Data?
This new tool certainly has the potential to have a big impact on how we prepare and deal with influenza outbreaks. Since COVID-19 is a novel virus, the lack of true historic data means it is difficult to train the AI system. The hope is that this promising AI application will spur more abundant and consistent public health data collection and help us better prepare for future pandemics and flu outbreaks. Unfortunately, even with improved data collection, we will still have to navigate the ongoing debate of whether public health law data should be used in AI systems. Ethical considerations in particular remain at the forefront of that debate, and it will remain to be seen whether such issues get resolved by the time data collection and predictive AI are ready for more general application.