Spatiotemporal variations in exposure: Chagas disease in Colombia as a case study

Abstract

Age-stratified serosurvey data are often used to understand spatiotemporal trends in disease incidence and exposure through estimating the Force-of-Infection (FoI). Typically, median or mean FoI estimates are used as the response variable in predictive models, often overlooking the uncertainty in estimated FoI values when fitting models and evaluating their predictive ability. To assess how this uncertainty impact predictions, we compared three approaches with three levels of uncertainty integration. We propose a performance indicator to assess how predictions reflect initial uncertainty.In Colombia, 76 serosurveys (1980–2014) conducted at municipality level provided age-stratified Chagas disease prevalence data. The yearly FoI was estimated at the serosurvey level using a time-varying catalytic model. Environmental, demographic and entomological predictors were used to fit and predict the FoI at municipality level from 1980 to 2010 across Colombia.A stratified bootstrap method was used to fit the models without temporal autocorrelation at the serosurvey level. The predictive ability of each model was evaluated to select the best-fit models within urban, rural and (Amerindian) indigenous settings. Model averaging, with the 10 best-fit models identified, was used to generate predictions.Our analysis shows a risk of overconfidence in model predictions when median estimates of FoI alone are used to fit and evaluate models, failing to account for uncertainty in FoI estimates. Our proposed methodology fully propagates uncertainty in the estimated FoI onto the generated predictions, providing realistic assessments of both central tendency and current uncertainty surrounding exposure to Chagas disease.

Linear and Machine Learning modelling for spatiotemporal disease predictions: Force-of-Infection of Chagas disease

“…The predictive R 2 values for the LM framework obtained, on average, for its 5 best-fitting models, were 77% and 70%, with %overlap of 54% and 39% for urban and rural settings, respectively [8]. For the ML frameworks, the MedFoI approach yielded substantially better predictive R 2 values (ranging between 90% and 98%), but the degree of overlap between the distributions of the FoI 'observations' and the predictions was substantially lower (19%-25%), reflective of a tighter distribution around the central estimates and thus indicating over-confidence in the predictions when using such a simple approach (i.e.…”

“…The 10 best models for each setting type (urban, rural and indigenous) were averaged and used to obtain FoI predictions. The LM framework has been fully described in [8].…”

“…The best ML models obtained were then compared with the LM framework previously developed [8] in terms of their performance indicators, predictions, and ability to propagate uncertainty. In addition to these aspects, the models' ability to deal with temporal and spatial correlation, as well as their different computational aspects entered the comparison.…”

“…Furthermore, when a non-constant (e.g. a yearly-varying) FoI is assumed, each serosurvey analysed becomes a temporal series at a certain location, requiring specific and computationally-intensive methods to be included into predictive models [8]. Machine Learning methods could represent a faster and more flexible framework to implement such models.…”

“…This paper aims to compare the performance of two ML methods, namely, BRT and RF, with a Linear Model (LM) framework we have previously developed [8] in their ability to predict the FoI of Chagas disease across space and time. We use detailed data from Colombia as a case study and describe the advantages and disadvantages of using such Machine Learning methods compared to Linear Model frameworks, specifically focussing on their ability to handle uncertainty on the response variable.…”