An Algorithm for Flu Outbreak Detection

Influenza is a constantly moving target for vaccination, and it always holds the potential to spread quickly. A new study found that predictive modeling may add a layer of protection against the common, but potentially dangerous, virus.

Using test reimbursement data made available from the Centers for Medicare and Medicaid Services (CMS), researchers — led by Rex Astles, PhD, coordinator at the Center for Disease Control (CDC) Division of Laboratory Systems in Atlanta, GA — an algorithm that can calculate predicted weekly volumes of both flu tests and diagnoses. The study, presented at the 69th American Association for Clinical Chemistry (AACC) Annual Scientific Meeting & Clinical Lab Expo in San Diego, CA, this week, used Medicare outpatient claims between 2007-2012.

Influenza episodes were defined by increases in predicted volume surpassing a certain standard deviation, followed by similar drops in volume. To compare detected flu episodes with those defined, the team used positive rapid flu test volumes from the CDC’s National Respiratory and Enteric Virus Surveillance System (NREVSS).

In the preliminary data, 67 influenza episodes were detected with the NREVSS. The team’s established algorithm wasn’t far from base — predicted weekly volumes of flu tests caught 64 (95.5%) of the episodes, and predicted weekly volumes of diagnoses caught 60 (89.5%). Additionally, the algorithm-based data was able to identify influenza active about 3 weeks earlier than NREVSS data been.

Though the data was closely accurate, it is only the results of data from 10 monitored states in the study. Researchers noted detection via CMS claims data would be effective in detecting influenza activity in "specific regions of the US."

Astles theorized the newfound approach could help augment the current standard of influenza tracking — not replace it.

“Our work showed that it is theoretically possible to use the simple fact that a test was ordered and performed as a means of detecting early respiratory disease activity,” Astles said.

The hope moving forward for the research is that it could be further implemented by larger healthcare systems with more vast administrative data sets, Astles said.

“This could be especially helpful for other diseases that have less effective surveillance systems,” Astles said. “Comprehensive surveillance has many purposes, only one of which is detection of apparent disease activity. This detection could be a signal that more intensive surveillance, including testing patient specimens to identify the circulating viral or bacterial strain, should be initiated.”

An abstract of the study, "B-085 - Evaluation of the Utility of CMS Claim Data for Early Detection of Increasing Influenza Activity," was made available on the AACC website, and the team released a press release regarding the work.

A version of this story originally appears in MD Magazine.Related Coverage:MIT Team Tests Object-Identifying Algorithm for the Visually ImpairedMicrosoft, Machine Learning, and Better Mosquito DefenseApplying Big Data to Disease Outbreaks