Real-time monitoring of local influenza can be a challenge because infections are spreading easily when people move and travel.
A study published in Nature Communications suggests that the ARGONet approach gives more accurate predictions than ARGO's earlier high-performance prognostic approach.
"Early and reliable methodologies for monitoring influenza activity at various sites can help public health workers alleviate the epidemic and improve communication with the public to raise awareness of potential risks," said Mauricio Santillana of the Computational Health Informatics Program (CHIP) at Boston Children's Hospital in the United States States.
The ARGONet approach uses machine learning and two robust models of influenza detection.
The first model, ARGO (AutoRegression with General Online Information), uses information from electronic health records, Google search, and influenza history.
The ARGO study itself outperformed Google's Flu Trends, a previous prediction system that ran from 2008 to 2015.
To improve accuracy, ARGONet adds a second model that draws from space-time flu specimens spreading in neighboring areas.
"It uses the fact that the presence of influenza in nearby localities can increase the risk of disease at the site," said Santillana, an assistant at Harvard Medical School.
The machine learning system was "trained" by providing flu predictions from both models as well as current flu data to help reduce predictions errors.
"The system continually evaluates the predictive power of each independent method and retrains how this information should be used to produce better flu estimates," Santillana said.
Scientists believe that their approach will provide the basis for "accurate public health" in infectious diseases.
"We believe our models will become more accurate over time as more online search volumes are collected and other health care providers have electronic cloud-based electronic health records," said Fred Lu, CHIP investigator.