Geographically weighted generalized Farrington algorithm for rapid outbreak detection over short data accumulation periods

Abstract

The demand for rapid surveillance and early detection of local outbreaks has been growing recently. The rapid surveillance can select timely and appropriate interventions toward controlling the spread of emerging infectious diseases, such as the coronavirus disease 2019 (COVID-19). The Farrington algorithm was originally proposed by Farrington et al (1996), extended by Noufaily et al (2012), and is commonly used to estimate excess death. However, one of the major challenges in implementing this algorithm is the lack of historical information required to train it, especially for emerging diseases. Without sufficient training data the estimation/prediction accuracy of this algorithm can suffer leading to poor outbreak detection. We propose a new statistical algorithm-the geographically weighted generalized Farrington (GWGF) algorithm-by incorporating both geographically varying and geographically invariant covariates, as well as geographical information to analyze time series count data sampled from a spatially correlated process for estimating excess death. The algorithm is a type of local quasi-likelihood-based regression with geographical weights and is designed to achieve a stable detection of outbreaks even when the number of time points is small. We validate the outbreak detection performance by using extensive numerical experiments and real-data analysis in Japan during COVID-19 pandemic.We show that the GWGF algorithm succeeds in improving recall without reducing the level of precision compared with the conventional Farrington algorithm.

Changes in marriage, divorce and births during the COVID-19 pandemic in Japan

“…To adjust for seasonality in the model, periods not included in the reference period are evenly divided into four subperiods, and each subperiod is encoded as binary dummy variables. The regression model is then given by 60 :…”

The convergence epidemic volatility index (cEVI) as an alternative early warning tool for identifying waves in an epidemic

“…Indeed, the positive and negative predictive values of the indices are impacted by the case definition and thus, their relative performance on identifying and retaining a warning, as observed in the empirical comparison of EVI approaches to the Farrigton's algorithm ( Table A1 ). We should note that the applied original Farrigton's algorithm has been further extended and alternatives may perform better when data are available on local outbreaks ( Yoneoka et al., 2021 ).…”

“…Such strategies can be supported by early warning tools/systems that provide timely indication and adoption of preventive measures. Early warning systems focus on one or many of the following aspects; 1) modeling of the epidemics’ seasonality, 2) identification of the link between meteorological parameters and pathogens ( Heffernan et al., 2004 ) and/or 3) spotting of spatial and temporal abnormalities in the expected number of cases ( Vega et al., 2013 ; Yoneoka et al., 2021 ).…”

Modeling and forecasting traffic flows with mobile phone big data in flooding risk areas to support a data-driven decision making

“…We use the Mean Absolute Percentage Error (MAPE) for the performance evaluation using the validation sample. Inspired by the famous algorithm for the early detection of outbreaks, proposed by Farrington et al (1996) (and developed by Yoneoka et al (2021)) and based on defining a cut-off for the intensity of a specific event, we then turn observed and predicted values into categories using quintiles representing different levels of exposure risk, and we evaluate predicting performance by means of the Hit-Rate (HR) measure for confusion matrices.…”