Characterizing Influenza surveillance systems performance: application of a Bayesian hierarchical statistical model to Hong Kong surveillance data

Abstract

Background: Infectious disease surveillance is a process the product of which reflects both actual disease trends and public awareness of the disease. Decisions made by patients, health care providers, and public health professionals about seeking and providing health care and about reporting cases to health authorities are all influenced by the information environment, which changes constantly. Biases are therefore imbedded in surveillance systems; these biases need to be characterized to provide better situational awareness for decision-making purposes. Our goal is to develop a statistical framework to characterize influenza surveillance systems, particularly their correlation with the information environment.

Methods: We identified Hong Kong influenza surveillance data systems covering healthcare providers, laboratories, daycare centers and residential care homes for the elderly. A Bayesian hierarchical statistical model was developed to examine the statistical relationships between the influenza surveillance data and the information environment represented by alerts from HealthMap and web queries from Google. Different models were fitted for non-pandemic and pandemic periods and model goodness-of-fit was assessed using common model selection procedures.

Results: Some surveillance systems - especially ad hoc systems developed in response to the pandemic flu outbreak - are more correlated with the information environment than others. General practitioner (percentage of influenza-like-illness related patient visits among all patient visits) and laboratory (percentage of specimen tested positive) seem to proportionally reflect the actual disease trends and are less representative of the information environment. Surveillance systems using influenza-specific code for reporting tend to reflect biases of both healthcare seekers and providers.

Conclusions: This study shows certain influenza surveillance systems are less correlated with the information environment than others, and therefore, might represent more reliable indicators of disease activity in future outbreaks. Although the patterns identified in this study might change in future outbreaks, the potential susceptibility of surveillance data is likely to persist in the future, and should be considered when interpreting surveillance data.

Figure 1

Selection flowcharts for Google search keywords.

Figure 2

Posterior distributions of coefficient β _j,t,m ( m  = 2,..,5; j  = 1,…,11) in “completeness” parameter θ _j,t (multiplier for the estimated incidence rate) as measure of correspondence between surveillance data and the information environment proxy data in the pandemic model. A) Coefficient for Google search index for seasonal flu terms; B) Coefficient for Google search index for symptoms; C) Coefficient for Google search index for flu medications; D) Coefficient for Google search index for non-flu terms.

Figure 3

Conceptual model for biases in influenza surveillance data.

Figure 4

Posterior distributions of “Completeness” coefficient β _j,t,m ( m  = 2,..,5; j  = 4,…,9) in “completeness” parameter θ _j,t (multiplier for the estimated incidence rate) as measure of correspondence between surveillance and the information environment proxy data for surveillance systems that are less correlated with the information environment during the pandemic period. β2 is the coefficient for Google search index of seasonal flu terms; β3 is the coefficient for Google search index for symptoms; β4 is the coefficient for Google search index of medications; β5 is the coefficient for Google search index of non-flu terms.

Figure 5

Posterior distributions of α _j,t,m (m = 2,..,7; j = 1,…,11) in “excess” parameter φ _j,t as measure of correspondence between surveillance and the information environment proxy data during the pandemic period. α2 is the coefficient for total number of alerts at HealthMap; α3 is the coefficient for the total number of unique alerts at HealthMap; α4 is the coefficient for number of healthcare facilities related alerts at HealthMap; α5 is the coefficient for %RSV from virological surveillance; α6 is the coefficient for Google search index of authority; α7 is the coefficient for Google search index of pandemic influenza terms.

Figure 6

Posterior distributions of α _j,t,m ( m  = 2,..,7; j  = 4,5,6) in “excess” parameter φ _j,t as measure of correspondence between surveillance and the information environment proxy data for surveillance systems that are less correlated with the information environment during the pandemic period. α2 is the coefficient for total number of alerts at HealthMap; α3 is the coefficient for the total number of unique alerts at HealthMap; α4 is the coefficient for number of healthcare facilities related alerts at HealthMap; α5 is the coefficient for %RSV from virological surveillance; α6 is the coefficient for Google search index of authority; α7 is the coefficient for Google search index of pandemic influenza terms.

Figure 7

Posterior distributions for the public awareness index and non-flu index coefficient as measure of correspondence between surveillance and the information environment proxy data during the non-pandemic model. A) ρ2: coeffcieint for non-flu index in the NP model; B) ρ3: coefficient for illness index in NP model; C) ρ4: coeffcieint for public awareness index in the NP model.

Figure 8

Comparison of posterior distributions for the public awareness index and non-flu index coefficient as measure of correspondence between surveillance and the information environment proxy data in 2007 and 2008. A) ρ2: coeffcieint for non-flu index in the NP model; B) ρ3: coeffcieint for public awareness index in the NP model.

Figure 9

Comparison of posterior distributions for non-flu index, illness index and public awareness index as measure of correspondence between surveillance and the information environment proxy data for hospitalization data series during the non-pandemic period. A) flu-HA; B) P&I-HA; C) P&I-HA(0-15 yr); D) P&I-HA(65+ yr).ρ2 is the coefficient for non-flu index, ρ3 is the coefficient for illness index, and ρ4 is the coefficient for public awareness index.