Probabilistic daily ILI syndromic surveillance with a spatio-temporal Bayesian hierarchical model

Abstract

Background: For daily syndromic surveillance to be effective, an efficient and sensible algorithm would be expected to detect aberrations in influenza illness, and alert public health workers prior to any impending epidemic. This detection or alert surely contains uncertainty, and thus should be evaluated with a proper probabilistic measure. However, traditional monitoring mechanisms simply provide a binary alert, failing to adequately address this uncertainty.

Methods and findings: Based on the Bayesian posterior probability of influenza-like illness (ILI) visits, the intensity of outbreak can be directly assessed. The numbers of daily emergency room ILI visits at five community hospitals in Taipei City during 2006-2007 were collected and fitted with a Bayesian hierarchical model containing meteorological factors such as temperature and vapor pressure, spatial interaction with conditional autoregressive structure, weekend and holiday effects, seasonality factors, and previous ILI visits. The proposed algorithm recommends an alert for action if the posterior probability is larger than 70%. External data from January to February of 2008 were retained for validation. The decision rule detects successfully the peak in the validation period. When comparing the posterior probability evaluation with the modified Cusum method, results show that the proposed method is able to detect the signals 1-2 days prior to the rise of ILI visits.

Conclusions: This Bayesian hierarchical model not only constitutes a dynamic surveillance system but also constructs a stochastic evaluation of the need to call for alert. The monitoring mechanism provides earlier detection as well as a complementary tool for current surveillance programs.

Figure 1. Spatial distribution of the five hospitals and corresponding buffers.

Figure 2. Temporal patterns of observed (oi) and expected ILI (ei) visits during 2006–2007.

Figure 3. Probability of alert at the stage of model fitting.

Top line is for posterior probability = 0.7, middle for 0.5, and bottom for 0.3.

Figure 4. Temporal chart of ILI visits, different alerts and associated factors.

(a) ILI counts and probability of alert during the validation stage based on weekly updated parameters. (b) The time series plots of ILI visits, weekly influenza isolation rate, temperature and vapor pressure, respectively.