Non-local validated parametrization of an agent-based model of local-scale Taenia solium transmission in North-West Peru

Abstract

The pork tapeworm, Taenia solium, is the cause of a preventable zoonotic disease, cysticercosis, affecting both pigs and humans. Continued endemic transmission of T. solium is a major contributor of epilepsy and other neurologic morbidity, and the source of important economic losses, in many rural areas of developing countries. Simulation modelling can play an important role in aiding the design and evaluation of strategies to control or even eliminate transmission of the parasite. In this paper, we present a new agent based model of local-scale T. solium transmission and a new, non-local, approach to the model calibration to fit model outputs to observed human taeniasis and pig cysticercosis prevalence simultaneously for several endemic villages. The model fully describes all relevant aspects of T. solium transmission, including the processes of pig and human infection, the spatial distribution of human and pig populations, the production of pork for human consumption, and the movement of humans and pigs in and out in several endemic villages of the northwest of Peru. Despite the high level of uncertainty associated with the empirical measurements of epidemiological data associated with T. solium, the non-local calibrated model parametrization reproduces the observed prevalences with an acceptable precision. It does so not only for the villages used to calibrate the model, but also for villages not included in the calibration process. This important finding demonstrates that the model, including its calibrated parametrization, can be successfully transferred within an endemic region. This will enable future studies to inform the design and optimization of T. solium control interventions in villages where the calibration may be prevented by the limited amount of empirical data, expanding the possible applications to a wider range of settings compared to previous models.

Fig 1. Simulated villages.

Geographical distribution of households, use of corrals and latrine ownership in the villages that were used in the development of the model. Panel a: control villages used for model calibration. Panel b: intervention villages used for model validation.

Fig 2. Pork distribution.

The pig that was slaughtered belongs to household 1. Pork portions are distributed to all household members, including those that recently ate pork. Other households are then prioritized according to their distance to household 1. Humans in these households that ate pork in the past week do not receive pork. If there are no portions left, humans from the furthest households do not receive pork either (in this case, household 4). Skin, muscle and bone portions may contain cysts while entrails never do.

Fig 3. Estimation of contamination levels (pig module).

The proglottids in the defecation site represented on the graph are typically shared between the pigs whose home range includes the site. The out-home range contamination term accounts for the occasional wandering of other pigs near the site. Disseminated eggs can be seen at the defecation site.

Fig 4. ABC posteriors distribution.

Posterior distributions of model calibration parameters for the two calibration setups: panel a: necro calibration setup (fourth ABC-SMC stage). Panel b: noNecro calibration setup (third ABC-SMC stage). Horizontal bold lines indicate the posterior distribution median, boxes cover the parameter values above the first quartile and below the third quartile and vertical lines show maximum and minimum values. Black dots show outliers.

Fig 5. Calibration results.

Proportion of infected pigs with different cyst burdens estimated using the necro and noNecro calibration setups as compared with observed data for control group villages 515, 566 and 567. The simulated curves are the product of the average over 750 runs using the minimum ABC distance parameter sets for both the necro and noNecro setups.

Fig 6. Cross-validation noNecro setup.

Cross-validation for 100 randomly selected simulations from the noNecro calibration setup, third stage. Estimated parameter values are compared to true values for the 6 calibration parameters with the identity line plotted as reference. The correlation coefficient between estimated and true parameter values is shown for each parameter together with the corresponding p value.

Fig 7. Cross-validation, simplified setup.

Cross-validation for 100 randomly selected simulations from the simplified calibration setup, third stage. Estimated parameter values are compared to true values for the 3 calibration parameters with the identity line plotted as a reference. The correlation coefficient between estimated and true parameter values is shown for each parameter together with the corresponding p value.