2021/04/20 by Tanvir Ferdousi, Lee W. Cohnstaedt, Ferdousi, Tanvir +3 · 2 citations
Mathematics · Medicine · #68T10 #COVID-19 epidemiological studies #Data-Driven Disease Surveillance #FOS: Computer and information sciences #G.3 #I.5.2 #I.5.4 #Machine Learning (cs.LG) #Mosquito-borne diseases and control
paper · pdf · doi:10.48550/arxiv.2104.10289
openalex publication_date 2021/04/20 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
The performance of data-driven prediction models depends on the availability\nof data samples for model training. A model that learns about dengue fever\nincidence in a population uses historical data from that corresponding\nlocation. Poor performance in prediction can result in places with inadequate\ndata. This work aims to enhance temporally limited dengue case data by\nmethodological addition of epidemically relevant data from nearby locations as\npredictors (features). A novel framework is presented for windowing incidence\ndata and computing time-shifted correlation-based metrics to quantify feature\nrelevance. The framework ranks incidence data of adjacent locations around a\ntarget location by combining the correlation metric with two other metrics:\nspatial distance and local prevalence. Recurrent neural network-based\nprediction models achieve up to 33.6% accuracy improvement on average using the\nproposed method compared to using training data from the target location only.\nThese models achieved mean absolute error (MAE) values as low as 0.128 on [0,1]\nnormalized incidence data for a municipality with the highest dengue prevalence\nin Brazil's Espirito Santo. When predicting cases aggregated over geographical\necoregions, the models achieved accuracy improvements up to 16.5%, using only\n6.5% of incidence data from ranked feature sets. The paper also includes two\ntechniques for windowing time series data: fixed-sized windows and outbreak\ndetection windows. Both of these techniques perform comparably, while the\nwindow detection method uses less data for computations. The framework\npresented in this paper is application-independent, and it could improve the\nperformances of prediction models where data from spatially adjacent locations\nare available.\n