Forecasting Dengue Incidence with the Chi-squared Automatic Interaction Detection Technique

Abstract

This paper presents the data-driven approach to forecast number of dengue infection in the two most populated provinces of Thailand: Bangkok and Nakhon Ratchasima. Our forecasting technique is the chi-squared automatic interaction detection (CHAID), which is a machine learning approach that adopts decision tree as a main data structure for building a model and applies a chi-square computation for node splitting. The CHAID algorithm is in a group of multivariate data analysis in that it takes more than one attributes to build a model. The input data used in our model building phase are from the remote sensing indices, oceanic sensors, and ground sources. Remote sensing index data are obtained from the National Oceanic and Atmospheric Administration (NOAA) of the United States. These indices are used as representatives to assess temperature, humidity and brightness in the atmosphere as well as the greenness conditions of plants over the specific Earth surface areas. The oceanic sensor data are ONI index computed by NOAA to announce El Nino or La Nina events, which are assumed to affect the growth rate of mosquito larva. From the model assessment, we found that the multivariate CHAID models in both provinces are more accurate than the univariate autoregressive integrated moving average (ARIMA) models that are currently used by the Thai public health practitioners. We also compare predictive performance of CHAID models against ARIMA and those obtained from other machine learning techniques. The CHAID models outperform the others. The induced CHAID model is thus considered efficient enough to be applied for predicting number of dengue infected patients.