While COVID-19 is currently at the forefront of public consciousness, it is far from the only infectious disease we need to worry about. For highly urbanized, tropical Singapore, dengue has been a perennial concern. 2020 saw a record high of over 38,000 cases, far surpassing the previous high of 22,318 cases in 2013.
Just as weather reports can help people plan their daily activities, disease forecasting models are helping public health officials prepare for potential dengue outbreaks. However, most existing dengue prediction models focus on predicting what the absolute number of cases will be in the next couple of weeks, rather than trying to detect a sharp, consecutive increase in cases that would suggest an impending outbreak.
“There are two problems with this approach,” explained Xiuju Fu, a Senior Scientist at A*STAR’s Institute of High Performance Computing (IHPC). “The first is that accuracy drops dramatically the farther ahead the model tries to see. Second, there is a lack of clear definition of the threshold, so such identification of dengue outbreaks can be quite subjective.”
Instead, Fu and her team developed a model that can detect dengue outbreaks at an early stage, buying time for public health officials to launch interventions and prepare resources. The key to their success was that they focused on broader differences between outbreak and non-outbreak situations rather than simply the number of cases.
The researchers first sought to determine what ‘normal’ years and periods would look like, using weekly dengue data from 2006 to 2011, years where numbers were relatively low. To establish the best baseline possible, they deliberately left out periods when an outbreak was in full swing. The resulting model was able to generate accurate forecasts a week in advance.
Borrowing the ‘control chart’ concept from manufacturing quality control processes, the researchers then input the differences between the forecasted value and the actual number of cases observed. If the difference deviates from the normal pattern, the model signals that an outbreak might be imminent. “Moreover, by using the proposed framework adaptively, the dynamic threshold for detecting outbreaks can be automatically determined, which makes the model intelligent and adaptive in changing environments,” Fu said.
Fu and her team are currently working on adding further detail such as spatiotemporal patterns to their predictions. This additional layer of information could help identify which areas are more prone to large-scale dengue outbreaks, helping officials launch more targeted—and thus less costly—proactive control measures.
The A*STAR-affiliated researchers contributing to this research are from the Institute of High Performance Computing (IHPC).