Extracting Temporal-Based Spatial Features in Imbalanced Data for Predicting Dengue Virus Transmission

Abstract

Since the movements of mosquito or human can potentially influence dengue virus transmission, recognizing location characteristics defined as spatial factors is necessary for predicting patient status. We proposed feature extraction that considers location characteristics through previous dengue cases and the high possibility of encounters between people with different backgrounds. The number of incoming populations, school buildings and population density was included as the location characteristics. Besides the information of the spatial factors, the number of dengue cases set within a particular time window was specified for virus transmission period. Our experiments obtained two datasets of dengue fever which were patient registry and location characteristics of Malang Regency. Manually recorded Registry Data only contained positive group data and not the negative group when the patients were healthy. Thus, the proposed extraction method also included the process of generating negative data from the existing positive data. Then, we preprocessed the data by cleaning, imputing, encoding, and merging, such that there were four features representing previous dengue cases and eight features describing location characteristics. The experiments demonstrated that by using some ranked features the prediction had a better accuracy of 78.7% compared to using all features. Temporal-based features displayed better performances, but the result was improved in the wider location where people met.