A Windowed Correlation-Based Feature Selection Method to Improve Time Series Prediction of Dengue Fever Cases

The performance of data-driven models depends on training samples. For accurately predicting dengue fever cases, historical incidence data are inadequate in many locations. This work aims to enhance temporally limited dengue case data by methodological addition of epidemically relevant case data from nearby locations as predictors (features). A novel framework is presented for windowing incidence data and computing time-shifted correlation-based metrics to quantify feature relevance. The framework ranks incidence data of adjacent locations around a target by combining metrics based on correlation, spatial distance, and local prevalence. Recurrent neural network models achieve up to 33.6% accuracy improvement on average using the proposed method. These models achieve mean absolute error (MAE) values as low as 0.128 on [0, 1] normalized incidence data for a municipality with the highest dengue prevalence in Brazil’s Espirito Santo. When predicting aggregate cases over geographical ecoregions, the models improve by 16.5%, using only 6.5% of ranked incidence data. This paper also presents two correlation window allocation methods: fixed-size and outbreak detection. Both perform comparably well, although the outbreak detection method uses less data for computations. The proposed framework is generalized, and it can be used to improve time-series predictions of many spatiotemporal datasets.