On Spatially Explicit Models of Epidemic and Endemic Cholera: The Haiti and Lake Kivu Case Studies.
Abstract:The first part of the Lecture deals with the predictive ability of mechanistic models for the Haitian cholera epidemic. Predictive models of epidemic cholera need to resolve at suitable aggregation levels spatial data pertaining to local communities, epidemiological records, hydrologic drivers, waterways, patterns of human mobility and proxies of exposure rates. A formal model comparison framework provides a quantitative assessment of the explanatory and predictive abilities of various model settings with different spatial aggregation levels. Intensive computations and objective model comparisons show that parsimonious spatially explicit models accounting for spatial connections have superior explanatory power than spatially disconnected ones for short-to intermediate calibration windows. In general, spatially connected models show better predictive ability than disconnected ones. We suggest limits and validity of the various approaches and discuss the pathway towards the development of case-specific predictive tools in the context of emergency management.
The second part deals with approaches suitable to describe patterns of endemic cholera. Cholera outbreaks have been reported in the Democratic Republic of the Congo since the 1970s. Here we employ a spatially explicit, inhomogeneous Markov chain model to describe cholera incidence in eight health zones on the shore of lake Kivu. Remotely sensed datasets of chlorophyll a concentration in the lake, precipitation and indices of global climate anomalies are used as environmental drivers in addition to baseline seasonality. The effect of human mobility is also modelled mechanistically. We test several models on a multi-year dataset of reported cholera cases. Fourteen models, accounting for different environmental drivers, are selected in calibration. Among these, the one accounting for seasonality, El Nino Southern Oscillation, precipitation and human mobility outperforms the others in cross-validation.