A Bayesian Network Approach to the Study of Historical Epidemiological
Total Page：16
File Type：pdf, Size：1020Kb
ResearchResearch A Bayesian network approach to the study of historical epidemiological databases: modelling meningitis outbreaks in the Niger A Beresniak,a E Bertherat,b W Perea,b G Soga,c R Souley,d D Duponte & S Hugonnetb Objective To develop a tool for evaluating the risk that an outbreak of meningitis will occur in a particular district of the Niger after outbreaks have been reported in other, specified districts of the country. Methods A Bayesian network was represented by a graph composed of 38 nodes (one for each district in the Niger) connected by arrows. In the graph, each node directly influenced each of the “child” nodes that lay at the ends of the arrows arising from that node, according to conditional probabilities. The probabilities between “influencing” and “influenced” districts were estimated by analysis of databases that held weekly records of meningitis outbreaks in the Niger between 1986 and 2005. For each week of interest, each district was given a Boolean- variable score of 1 (if meningitis incidence in the district reached an epidemic threshold in that week) or 0. Findings The Bayesian network approach provided important and original information, allowing the identification of the districts that influence meningitis risk in other districts (and the districts that are influenced by any particular district) and the evaluation of the level of influence between each pair of districts. Conclusion Bayesian networks offer a promising approach to understanding the dynamics of epidemics, estimating the risk of outbreaks in particular areas and allowing control interventions to be targeted at high-risk areas. Introduction zation (WHO) consists of the early detection of epidemics, the treatment of cases with antibiotics, and mass vaccination to Throughout the African “meningitis belt”, epidemics of me- halt the outbreak (if possible, within 4–6 weeks of the epidemic ningococcal disease have been reported since the disease was threshold being reached).8,14,15 The early detection of meningitis first described, in 1912.1 Every year, western African countries epidemics is based mainly on the observation of the trends in within the Sahelo–Sudanian band experience major outbreaks weekly incidence.8,12,14,16–18 of meningococcal meningitis, each of which can affect up to The effectiveness of mass vaccination as a strategy for the 200 000 people, most of them young children.2 Burkina Faso, control of meningitis epidemics has been questioned.19 Given Mali and the Niger, for example, are regularly hit by meningi- the sporadic nature of the outbreaks, the optimal use of vaccines tis epidemics.3–7 Meningitis is characterized by high levels of to control both short-term epidemic and endemic meningococ- seasonal endemicity, with large epidemics of meningococcal cal disease has been the subject of much debate.20 In particular, meningitis occurring cyclically.8 Each such epidemic typi- the results of several studies in Africa have shown that vaccina- cally starts in January and ends in late May.2 The three main tion during outbreak situations is suboptimal, mainly because pathogens causing bacterial meningitis in Africa are Neisseria populations in resource-poor areas cannot be immunized rap- meningitidis, Haemophilus influenzae type b and Streptococcus idly enough.20 In addition, although rapid laboratory diagnosis pneumoniae.4,9,10 is an essential component in the surveillance of meningococcal As a major cause of morbidity and mortality in sub-Sa- epidemics, as it allows decision-makers to select the most ap- haran Africa, meningitis merits specific control measures.11–13 propriate vaccine for mass vaccination,21–23 the resource-poor Burkina Faso, Mali and the Niger have each already established countries most affected by such epidemics struggle to achieve a broad surveillance network for collecting data on cases of such diagnosis.21 A new conjugate vaccine that protects against the disease.4,6,11,12 Since 1986, peripheral health centres in each serogroup A meningitis (the form involved in most outbreaks of of these countries have routinely collected data on suspected meningococcal disease in Africa) was approved by the United cases of meningitis and reported each such case to health States Food and Drug Administration in February 2010. Con- districts for subsequent analysis at the national health minis- jugate vaccines prevent both carriage and the transmission of try. The information collected in this surveillance now forms bacteria from person to person. important epidemiological databases that could improve our Extensive descriptive analyses that have been performed understanding of the nature of meningitis epidemics and allow on meningitis surveillance databases have provided important high-risk areas to be identified. Current strategies for control- information on epidemic cycles, seasonality and the correla- ling meningitis epidemics in sub-Saharan Africa are based on tions between morbidity and certain co-factors, such as climatic epidemiological, immunological and logistical considerations. parameters.2,24–26 While the seasonal and spatial patterns of The strategy currently advocated by the World Health Organi- the disease appear to be linked to climate, the mechanisms a Data Mining International, Route de l’Aeroport 29–31, CP 221, 1215 Geneva 15, Switzerland. b Epidemic and Pandemic Alert and Response, World Health Organization, Geneva, Switzerland. c World Health Organization, Niamey, Niger. d Department of Statistics, Surveillance and Response to Epidemics, Ministry of Health, Niamey, Niger. e Data Mining America, Montreal, Canada. Correspondence to A Beresniak (e-mail: [email protected]). (Submitted: 10 January 2011 – Revised version received: 5 December 2011 – Accepted: 5 December 2011 – Published online: 20 January 2012 ) 412 Bull World Health Organ 2012;90:412–417A | doi:10.2471/BLT.11.086009 Research A Beresniak et al. Bayesian modelling of meningitis outbreaks in the Niger Fig. 1. Simple Bayesian network node of a Bayesian network has an associ- for “epidemic alert”, which could be used composed of nodes, arrows and ated conditional probability table. In an for the early detection of a meningitis conditional probabilities for epidemiological analysis, the nodes might epidemic, are lower (at least five cases the occurrence of a meningitis represent the districts of a country and the per 100 000 inhabitants and at least two epidemic in each of a country’s arrows and conditional probability table cases per 100 000, respectively). five hypothetical districts might show how a disease outbreak in Two types of analysis were then one district is linked to the probability of a performed, one (the “first-level” analy- disease outbreak in another district. In the sis) taking no account of the time taken District A District E example shown (Fig. 1), if district A expe- for an epidemic in one district to influ- Outbreak Outbreak rienced an outbreak, the probabilities that ence the development of an epidemic districts B and C experienced outbreaks elsewhere (ignoring the timing of the 15% 22% 85% would be 85% and 15%, respectively. epidemics) and the other (the “sec- Since January 1986, WHO has su- ond-level” analysis) assuming that an pervised the weekly collection of data epidemic in one district would influence District B 91% District C on meningitis incidence in each district the development of an epidemic else- Outbreak Outbreak of the Niger and these data now form a where 1, 2 or 3 years later. A maximum substantial historical database. The data lag of 3 years was explored because this 36% District D recorded include district name and popu- was both the median period in the cycles Outbreak lation, week number and the reported of meningitis outbreaks detected in the numbers of suspected cases of meningitis historical database and the estimated and of deaths attributed to meningitis in duration of protection resulting from a responsible for these patterns have still that district and week. For the present vaccination programme. to be elucidated.2,25 study, the data from the database for the The results of the “first-level” analy- In the present study, surveillance 14 calendar years from January 1986 to sis, performed with version 1.0 of the data and a modelling method based on December 1999 were used to construct Discoverer software package (Bayes- Bayesian networks were used to explore a Bayesian network, with a node for ware, Milton Keynes, United Kingdom how meningitis incidence in a district of each of the 38 districts in the Niger. The of Great Britain and Northern Ireland), the Niger was influenced by, or influenced, weekly incidence thresholds set by WHO were used to create a graph of the Bayes- the incidence in any other district. The aim for epidemic meningitis (i.e. at least 10 ian network and included the relevant was to develop a method for the optimiza- cases per 100 000 inhabitants in a district conditional probability table, showing tion of epidemic alerts and the spatial and with a population of at least 30 000 or at the probabilities of meningitis out- temporal targeting of immunizations and least five cases per 100 000 in a district breaks. Because of the limited capability other interventions for the management of with a smaller population) were used of the Discoverer software, the analysis meningitis in the Niger, with the ultimate to give each district a weekly score of was restricted to the first 14 years of the goal of preventing meningitis epidemics 1 (if an epidemic had occurred) or 0 (if meningitis database for the Niger. The in the country. no outbreak had occurred). The corre- “minimum description length” learn- sponding weekly incidence thresholds ing algorithm was used, with network Methods A Bayesian network consists of a graphi- cal model showing the probabilistic Fig. 2. Bayesian network graph showing how each of the Niger’s 38 districts is influenced by, and/or influences, the probability of a meningitis outbreak in at relationships between one or more vari- least one other district ables of interest.