Spatial epidemiology and predictive modelling of Rift valley fever in Garissa county, Kenya

Author: Nanyingi, Mark Opiyo

Supervisors: Gerald M.; Stephen G.; Benard Bett and Samuel T. M.

Rift Valley fever (RVF) is an emerging arboviral zoonotic disease of domesticated livestock that causes disease in humans and has significant socio-economic impacts. In domestic ruminants, RVF causes abortions in females and high mortality rates in newborns. In humans, the disease begins with mild to acute febrile illness, and may progress to a severe haemorrhagic syndrome. Kenya has experienced two major outbreaks in the last 20 years; the 1997 outbreak that resulted in 89,000 human infections and 170 human deaths reported in Garissa and Southern Somali, and the 2006-2007 outbreak that resulted in livestock deaths estimated at US$10 million in economic loss, and additional production losses related to abortions and decline in milk production estimated at US$77,000. Outbreaks of RVF in Garissa have been associated with high rainfall and flooding due to drastic climate variability, and the introduction of the virus associated with livestock movement and trade with neighbouring countries have enhanced RVF endemicity. The lack of preparedness of local communities which includes lack of knowledge on cause, risks and symptoms, poor attitude and practices in undertaking effective management increases their vulnerability to the disease outbreaks. Limited resources in health and livestock sectors have hampered routine serological monitoring of RVF virus activity to guide mitigation strategies. The lack of adaptable tools for risk communication of impending disease outbreaks lead to delayed response and control of RVF outbreaks in Garissa. To address the above challenges a study was conducted in Garissa with the goal to investigate the spatial distribution and risk of Rift Valley Fever in Garissa using modelbased approaches in prediction of outbreaks in response to climate variability. This was specifically achieved by evaluation of the knowledge, attitude and practices (KAP) of communities living in Garissa County on RVF outbreaks; determination of RVF seroprevalence in cattle, sheep, and goats during an interepidemic period and investigation of the spatial risk distribution and development of risk maps as part of an early warning system (EWS) tools. A systematic review of existing literature highlighting RVF epidemiology with a local and global perspective emphasising on descriptive epidemiology, spatial and temporal distribution, molecular epidemiology, serological surveillance and control measures was done at the beginning of the study. This was followed by a two-week reconnaissance survey that was conducted in 2012 to identify the study area, select the target human and animal populations which was crucial in designing the study and acquisition of secondary datasets. Participatory epidemiology was used where 275 local participants were selected and engaged through face to face interviews, key informant interviews and focus group discussions. Descriptive statistics were used to characterise levels of knowledge, attitudes and practices while logistic regression analyses were conducted to identify predictors of knowledge and practices. In July 2013, a cross-sectional survey was conducted where 370 ruminants were sampled from eight RVF prone areas of Garissa to determine the seroprevalence of RVF. Rift Valley fever virus (RVFV) antibodies were detected using a multispecies competitive enzyme-linked immunosorbent assay (c-ELISA) for specific detection of RVFV IgG antibodies. Mixed logistic regression models were used to determine the association between RVF seropositivity and species, sex, age, and location of the animals. In order to determine the influence of environmental, climatic and demographic drivers on the spatial distribution and predict the risk in of RVF occurrence in Garissa, the relationship between these predictor variables and seroprevalence of the ruminants was analyzed using species distribution modelling (SDM) approaches. A robust machine learning technique (Boosted Regression Trees (BRT) and a Bayesian hierarchal geostatistical model (Integrated Nested Laplace Approximation (INLA) were used to analyse spatial distribution of Rift Valley fever seropositivity. The predictive power of all models was evaluated, validated to determine the level of agreement in accurately predicting the spatial pattern and risk of Rift Valley fever. Two hundred and seventy-five people participated in the KAP survey, all (100%) of the 214 males (78%) and 61 females (22%) had heard of RVF, majority of males (61.7%) had high knowledge of RVF, while 82% of females had low knowledge. Results from the logistic regression analyses show that males had a fourfold likelihood (Odds Ratio- OR= 4.25, 95%CI 1.99-9.06) of being knowledgeable about RVF than females. Stormy abortions were identified as the most recognizable clinical sign in animals by 71.6% of the participants, while 50.2% of participants indicated high fever as common sign in humans. Increasing age and conversely lack of formal education were strongly associated with high levels of knowledge. The relationship between the overall knowledge of RVF and the respondents gender was statistically significant (χ2=36.23, df=1, p-value <0.001). RVF was controlled mainly through livestock vaccinations and avoiding consumption of animal products. The serological survey detected a high overall seroprevalence of 27.6% (95% CI [23– 32.1]) of RVF in 370 (271 goats, 87 sheep, and 12 cattle) ruminants. Sheep, cattle, and goats had seroprevalences of 32.2% (95% CI [20.6–31]), 33.3% (95% CI [6.7–60]), and 25.8% (95% CI [22.4–42]), respectively. Seropositivity in male species was 31.8% (95% CI [22.2–31.8]), whereas that of females was 27% (95% CI [18.1–45.6]). Animals greater than 1year-old had an 18-fold likelihood to be seropositive than animals less than 1 year, OR 18.24 (95% CI 5.26-116.4), xvii The spatial model, BRT predicted approximately 16,810 km2 of very high risk (predicted probability of 0.70), which is 70% of the total area of Garissa County. High precipitation (35%), high human and livestock density (27%) and elevated temperatures (17 %) were the most important variables to predict high risk for RVF occurrence. The BRT model predicted the highest risk (>0.5-1.0) in the north-western parts of Garissa, areas adjacent to the Tana River with widespread foci of medium to low risk (<0.3), central parts and around perennial water bodies. The predictive performance of BRT model was high AUC of ROC score of (0.7 ±0.001 s.d). INLA predicted overall medium risk (<0.3) (mean of the spatial component), with small areas (those close to observed data locations) having very high risk (>0.5-1.0). The predictive performance of INLA was found to be very high with AUC of ROC score of 0.9 ±0.001 s.d). There was a significant positive correlation and good model agreement between INLA/RF (global correlation, r =0.44) in predicting the serologic status of RVF in Garissa. The community had high knowledge on RVF symptoms and transmission and recognized vaccination as the main prevention strategy. Assessment of knowledge and practices of a population is a first step for planning public health and veterinary intervention before and during RVF outbreaks in Garissa. Despite the overall high seroprevalence of RVF in all ruminants, older animals had high likelihood of being seropositive due to prolonged exposure. xviii The endemicity of RVF in Garissa may be enhanced by movement of live (infected) animals to Garissa livestock market from Somalia and may be responsible for reintroduction of disease. Seropositive animals tended to aggregate near wetlands indicating spatial dependency of these cases in north western, central and southern parts of Garissa. This is the first study in Kenya that has used logistic regression and geostatistical models to investigate predictive factors associated with RVF prevalence in livestock, using actual serological data. It has provided predictive model-based risk maps for Garissa, at a high spatial resolution by exploring the underlying spatial processes and displayed high risk incidence areas. The generated predictive risk maps might be suggestive for areas to be targeted for RVF sentinel surveillance and intervention, this is useful for policy decision-makers to prioritize intervention areas for cost-effective and optimal resources allocation for RVF prevention and control. The findings of this study contribute to the overall understanding of the risk factors that predispose communities and their livestock to potential RVF outbreaks. It also provides baseline serological status of the disease and precisely predicts where the next likely outbreaks will occur which is very useful in assisting strategic, targeted and costeffective preparedness and implementation of control measures.