Predicting geographic variation in cutaneous leishmaniasis, Colombia

Abstract

Approximately 6,000 cases of cutaneous leishmaniasis are reported annually in Colombia, a greater than twofold increase since the 1980s. Such reports certainly underestimate true incidence, and their geographic distribution is likely biased by local health service effectiveness. We investigated how well freely available environmental data explain the distribution of cases among 1,079 municipalities. For each municipality, a unique predictive logistic regression model was derived from the association among remaining municipalities between elevation, land cover (preclassified maps derived from satellite images), or both, and the odds of at least one case being reported. Land cover had greater predictive power than elevation; using both datasets improved accuracy. Fitting separate models to different ecologic zones, reflecting transmission cycle diversity, enhanced the accuracy of predictions. We derived measures that can be directly related to disease control decisions and show how results can vary, depending on the threshold selected for predicting a disease-positive municipality. The results identify areas where disease is most likely to be underreported.

Figure 1

Distribution of a) ecoepidemologic zones, b) elevation, and c) vegetation types in Colombia.

Figure 2

Incidence of American cutaneous leishmaniasis per rural population reported in Colombia by year, 1980–2002 (data from Ministerio de Salud, Colombia).

Figure 3

Geographic distribution of American cutaneous leishmaniasis incidence by municipality, 1994

Figure 4

Performance of whole country model versus combination of zonal models. A. Receiver operator curve. Black line, single model for all Colombia (area under the curve [AUC] = 72.4%); gray line, combination of zonal models (AUC = 84.4%). Diagonal line indicates success expected on the basis of chance (AUC = 50%). B. κ value, representing skill at discriminating positive and negative municipalities, above the level expected on the basis of chance. Black line, single model for all Colombia; gray line, combination of zonal models. The probability threshold is the value on the continuous scale of predicted probability of transmission that is used as the cut-off for conversion into a categorical prediction of presence versus absence.

Figure 5

A. Sensitivity (solid line) and specificity (broken line) across the range of threshold probabilities for predicting an endemic municipality. B. Positive predictive value (solid line) and negative predictive value (broken line) across the range of threshold probabilities for predicting a positive municipality. The probability threshold is the value on the continuous scale of predicted probability of transmission that is used as the cut-off for conversion into a categorical prediction of presence versus absence.

Figure 6

Predicted risk map for probability of transmission, based on the combination of the regional models.

Figure 7

Agreement between predictions and observations. Light blue, correct positive prediction; light red, correct negative; dark blue, false positive; dark red, false negative.