Predicting high dengue incidence in municipalities of Brazil using path signatures

TL;DR

This paper presents a novel method using path signatures and lasso regression to predict high dengue incidence in Brazilian municipalities, achieving high specificity and sensitivity, and demonstrating potential for resource optimization in disease surveillance.

Contribution

It introduces a new framework combining path signatures with epidemiological and environmental data for improved dengue risk prediction.

Findings

Sensitivity reached 75%

Specificity ranged from 75% to 100%

Optimal predictions achieved after 35 weeks of data

Abstract

Predicting whether to expect a high incidence of infectious diseases is critical for health surveillance. In the epidemiology of dengue, environmental conditions can significantly impact the transmission of the virus. Utilizing epidemiological indicators alongside environmental variables can enhance predictions of dengue incidence risk. This study analyzed a dataset of weekly case numbers, temperature, and humidity across Brazilian municipalities to forecast the risk of high dengue incidence using data from 2014 to 2023. The framework involved constructing path signatures and applying lasso regression for binary outcomes. Sensitivity reached 75%, while specificity was extremely high, ranging from 75% to 100%. The best performance was observed with information gathered after 35 weeks of observations using data augmentation via embedding techniques. The use of path signatures effectively captures the stream of information given by epidemiological and climate variables that influence dengue transmission. This framework could be instrumental in optimizing resources to predict high dengue risk in municipalities in Brazil and other countries after learning these country patterns.