Construction and evaluation of two computational models for predicting the incidence of influenza in Nagasaki Prefecture, Japan

Fei He,Zhi-Jian Hu,Kiyoshi Aoyagi,Lin Cai,Guo-Xi Cai,Wen-Chang Zhang

doi:10.1038/s41598-017-07475-3

Abstract

It remains challenging to forecast local, seasonal outbreaks of influenza. The goal of this study was to construct a computational model for predicting influenza incidence. We built two computational models including an Autoregressive Distributed Lag (ARDL) model and a hybrid model integrating ARDL with a Generalized Regression Neural Network (GRNN), to assess meteorological factors associated with temporal trends in influenza incidence. The modelling and forecasting performance of these two models were compared using observations collected between 2006 and 2015 in Nagasaki Prefecture, Japan. In both the training and forecasting stages, the hybrid model showed lower error rates, including a lower residual mean square error (RMSE) and mean absolute error (MAE) than the ARDL model. The lag of log-incidence, weekly average barometric pressure, and weekly average of air temperature were 4, 1, and 3, respectively in the ARDL model. The ARDL-GRNN hybrid model can serve as a tool to better understand the characteristics of influenza epidemic, and facilitate their prevention and control.

Highlights

Influenza virus is the most common cause of acute respiratory illness[1]
In 2014 Lazer et al reported that the number of influenza-like outpatients reported by nation-wide CDC laboratory monitoring in 2013 was twice that predicted by Google Flu Trends (GFT) prediction service compared by using CDC, which came to the conclusion from laboratory-based monitoring reports across America, indicating the failure of the GFT was recognized as a defective model[6], because of big data hubris and change of algorithm dynamics
We validated the capacity of these two models to model and forecast influenza incidence, using influenza incidence data collected in Nagasaki prefecture between 2006 and 2015 by the Japanese infectious disease surveillance systems

Summary

Introduction

Influenza virus is the most common cause of acute respiratory illness[1]. influenza infection is usually self-limiting, it affects all age groups around the world and cause severe complications in high-risk individuals such as children, the elderly and those with chronic medical conditions. In Japan, influenza was designated in 1947 as a notifiable disease under the Japanese Communicable Disease Prevention Law, and systematic surveillance of influenza/influenza-like illness started in 1981. The relative high-frequency key words would always be used by GFT, search engine algorithms would produce adverse effect on forecast results of GFT. We validated the capacity of these two models to model and forecast influenza incidence, using influenza incidence data collected in Nagasaki prefecture between 2006 and 2015 by the Japanese infectious disease surveillance systems (http://www.nih.go.jp/niid/en/idsc-e.html). The performance of these two models was compared to identify the best influenza forecasting model

Objectives

Methods

Results

Conclusion