Early detection of influenza outbreaks: an application of a Bayesian online change point detection algorithm with optimal hyperparameter estimation using the CDC Influenza-Like Illness Surveillance Network (ILINet) data

Introduction Each year, the incidence of influenza (flu) and its financial costs are substantial in the United States (US). The Centers for Disease Control and Prevention (CDC) indicates that approximately 25 million people in the US were infected with influenza during the 2015-2016 flu season, leading to 11 million flu-related medical visits, and 12,000 flu-associated deaths. Although flu outbreaks occur every year, the timing and severity of these outbreaks vary from year to year. A critical component in averting the spread of the flu and its adverse consequences is early detection of imminent flu outbreaks. The earlier the detection, the more time there is to implement proactive prevention strategies against the spread of the disease. However, under the current gold standard for flu surveillance, the US Outpatient Influenza-Like Illness Surveillance Network (ILINet) conducted by the CDC, flu activity is estimated and monitored based on clinical and laboratory data. As such, there is always a delay of up to three weeks between the occurrence of the outbreak and dissemination of this information. Thus, there is an urgent need for improving and strengthening the flu surveillance system to provide timely outbreak information for guiding public health decisions that seek to prevent and control the disease. Purpose To test the feasibility for early detection of imminent flu outbreaks by applying a Bayesian online change point detection (BOCPD) algorithm with optimized hyperparameter estimation to the CDC’s ILINet data. Method CDC ILINet data from the week of 1/16/2010 through 4/29/2017 (N = 380 weeks) were used in the analysis. The CDC ILINet data consist of weekly number of people seeking medical attention with symptoms of influenza-like illness (ILI). Change points were detected using the BOCPD algorithm with a 1-year (52 weeks) rolling window. That is, instead of using a constant set of hyperparameters for the machine learning process, new hyperparameters were used every week when detecting change points, where the new hyperparameters were estimated using the data from the previous 52 weeks. CDC declares flu outbreaks using ILINet data when the percent of ILI during that week exceeds a predetermined threshold. For each season, the first change point that satisfied the following conditions was considered informative in early detection of the subsequent outbreak: (1) its percent of ILI visit was higher than that of the last change point, (2) its percent of ILI visit was less than the predetermined CDC threshold for an outbreak, and (3) the relative change between its percent of ILI and the CDC threshold was less than 50%. Results Except for the 2011-2012 flu season, we were able to detect the imminent outbreak, on average, 6 weeks prior to the actual outbreak. Conclusion Results suggest that the BOCPD algorithm may be effective in detecting flu outbreaks weeks prior to the start of the outbreak.