The Global Burden of Meningitis in Children: Challenges with Interpreting Global Health Estimates

The World Health Organization (WHO) has developed a global roadmap to defeat meningitis by 2030. To advocate for and track progress of the roadmap, the burden of meningitis as a syndrome and by pathogen must be accurately defined. Three major global health models estimating meningitis mortality as a syndrome and/or by causative pathogen were identified and compared for the baseline year 2015. Two models, (1) the WHO and the Johns Hopkins Bloomberg School of Public Health’s Maternal and Child Epidemiology Estimation (MCEE) group’s Child Mortality Estimation (WHO-MCEE) and (2) the Institute for Health Metrics and Evaluation (IHME) Global Burden of Disease Study (GBD 2017), identified meningitis, encephalitis and neonatal sepsis, collectively, to be the second and third largest infectious killers of children under five years, respectively. Global meningitis/encephalitis and neonatal sepsis mortality estimates differed more substantially between models than mortality estimates for selected infectious causes of death and all causes of death combined. Estimates at national level and by pathogen also differed markedly between models. Aligning modelled estimates with additional data sources, such as national or sentinel surveillance, could more accurately define the global burden of meningitis and help track progress against the WHO roadmap.

Pathogen specific meningitis mortality and incidence modelling methods 4 GBD 2017 4 MCEE/JHSPH 5 Tables and Figures   Table S1: ICD10 codes mapped to meningitis, encephalitis and neonatal sepsis according to model 7 Table S2: Definitions of IHME's GBD 2017 data quality star rating 12     Figure S1: Quality of underlying cause of death data and modelling methods used to generate death estimates according to model. 17 Mortality modelling methods -Meningitis, encephalitis and neonatal sepsis This section provides a brief overview on how the different models derive meningitis, encephalitis and neonatal sepsis mortality estimates. However, full methodology is provided by the modellers elsewhere.
Both GBD 2017 and MCEE 2000-2017 models used cause of death data from vital registration (VR) systems, sample registration systems (SR) and verbal autopsy (VA) studies to assign causes of death ensuring that the total number of deaths matches other estimates for the age-specific all-cause mortality. This involves the generation of data where data are The global burden of meningitis in children: Challenges with interpreting global health estimates Supplementary Appendix 1 incomplete or completely missing. The models also try to correct for poor quality cause of death data.
A description of each model is outlined below.

GBD 2017
A core step in the GBD 2017 estimation process was the creation of a cause of death (CoD) database where cause of death data obtained from individual countries is mapped to the GBD cause list of 282 diseases and split into age and sex categories [3].
The International Classification of Diseases ICD10 codes used to map to the GBD cause categories "meningitis", "encephalitis" and "neonatal sepsis and other neonatal infections" are outlined in Table S1.
Some causes of death were not considered specific enough to be mapped to a particular GBD cause of death category, could not be the underlying cause of death (e.g. senility) or had been assigned to the immediate or intermediate cause of death rather than the underlying cause (e.g. heart failure) and were therefore considered to be garbage codes. A statistical process was used to redistribute garbage codes to other causes of death and smooth out unrealistic data points.
For GBD 2017 meningitis death estimates, a total of 19,331 vital registration (VR) data points, 1,470 verbal autopsy (VA) data points, 793 sample registration (SR) data points and 546 surveillance data points were used [3]. A data point represents cause of death data in an individual location for a specific year. For example, VR data from a specific location for the years 1990-2017 inclusive would equate to 28 data points.
VR country-years with data less than 50% complete were dropped and country data with completeness between 50 -69% were flagged as non-representative in the CoD database. Raw data points from VR and SR were adjusted using death distribution methods that assess the completeness of death recording relative to census recording. Raw age specific mortality was then divided by estimated completeness to account for under ascertainment in VR and SR data.
To address cause of death data that is incomplete or not available for many locations IHME used Cause of Death Ensemble Models (CODEm) to fill in the gaps in the data by drawing on data from countries with more complete data, similar characteristics and geography. Different CODEm were used to estimate meningitis deaths according to sex and in the 0 days to 4 years age group compared to 5 years and older. Additionally different CODEm models were used to estimate meningitis deaths from data rich locations (in countries with 4star or greater rated VR systems) compared to the global model which includes all countries and data and includes those countries and territories where data was less reliable or where there was no data at all [2]. Definitions of the IHME data quality star rating is outlined in Table S2. In locations with data, the in-country data is heavily weighted, and data from the region and super region has a minor influence. In locations without data, estimates are informed by covariates and by data from the region and super region.
Location-level covariates used in CODEm models are outlined in Table S3.
Estimates generated from the CODEm models were then combined with other cause of death estimates ensuring that the sum matched the total all-cause mortality envelope for The global burden of meningitis in children: Challenges with interpreting global health estimates Supplementary Appendix each age group, sex, location and year. The all-cause mortality envelope is generated using a combination of surveys, censuses and vital registration data.

GBD 2017 encephalitis death estimates
Deaths from encephalitis were modelled using CODEm. The covariates used are outlined in Table S3. A total of 19,028 vital registration (VR) data points, 395 verbal autopsy (VA) data points and 793 sample registration (SR) data points were used [3].
GBD 2017 neonatal sepsis death estimates Deaths from neonatal sepsis were modelled in children under 5 years using CODEm in four separate age groups: Early neonatal period, late neonatal period, post neonatal period and 1-4 years. The covariates used are outlined in Table S3. A total of 18,175 vital registration (VR) data points, 165 verbal autopsy (VA) data points and 791 sample registration (SR) data points were used [3]. Modellers excluded the majority of verbal autopsy data (apart from in India) in the estimation of neonatal sepsis deaths because validation studies indicate that verbal autopsy methods are less accurate or defining cause of death in this age group.

WHO-MCEE 2000-2017
WHO-MCEE calculated cause of death fractions according to a predefined cause list (Table  S5). These cause of death fractions were then applied to neonatal and 1-59 month all-cause estimates produced by UN-IGME. The underlying data used by UN-IGME for calculating mortality rate and deaths comes from surveys, censuses and vital registration data.
Three methods were used to calculate cause of death fractions for a country depending on the quality of the cause of death data available for that country and its mortality setting.
VR data -For countries with high quality data covering >80% of the population VR data was used directly to estimate cause of death fractions attributed to the cause categories described in Table S5. Data was defined as high quality if countries had reported at least five years of data to WHO with an average usability over this period of equal to or over 80%. Usability is calculated as completeness of data multiplied by the proportion of registered deaths that are assigned a meaningful cause [27,49]. The ICD10 codes used to map to meningitis, encephalitis and neonatal sepsis are provided in Table S1.
VRMCM -In low mortality countries (<35 deaths/1000 live births 2000-2010) data from the countries with high quality VR were used to fit a multinomial logistic regression model, called the vital registration multi-cause model (VRMCM), using the covariates outlined in Tables 3  and 4. Cause of death fractions were attributed to the cause categories described in Table  S6.
VAMCM -In high mortality countries (>35 deaths/1000 live births 2000-2010) the cause distribution was estimated using a multinomial model applied to verbal autopsy data. The verbal autopsy multi-cause model (VAMCM) was a multinomial logistic regression model fitted using VA data from 119 studies in 39 countries using the covariates outlined in Tables  3 and 4. Cause of death fractions were attributed to the cause categories described in Table  S6.
Causes of death within the early neonatal (0-6 days), late neonatal (7-28 days) and post neonatal period (0-11 months) were modelled separately from each other because the cause of death distributions can differ significantly within these age groups. Different country specific covariates were used to derive the predictions based on age, model type and cause (Table S4). Despite the early and late neonatal period being modelled separately, only neonatal and post neonatal cause of death estimates are published. Post -hoc adjustment covariates were also used to account for meningitis deaths averted by PCV and Hib vaccines. The adjustments take into account serotype coverage of the vaccine in the case of PCV, vaccine coverage and vaccine effectiveness.
In the MCEE 2000-2017 estimation round neonatal meningitis was estimated separately from neonatal sepsis for the first time. These causes were estimated separately for the first time in their latest modelling round by using the ratio of neonatal meningitis and neonatal sepsis deaths derived from IHME estimates. The six recognised direct causes of neonatal deaths identifiable by verbal autopsy are: (1) serious infection (including sepsis, pneumonia and meningitis), (2) birth asphyxia, (3) prematurity, (4) tetanus, (5) congenital malformation and (6)  Pathogen specific meningitis mortality and incidence modelling methods

GBD 2017
Pathogen specific mortality estimates GBD 2017 assigned overall deaths from meningitis as predicted using CODEm into pathogenic causes using a set of proportional models in DisMod-MR 2.1. Proportions were informed using vital registration death data coded down to cause-level. The meningococcal meningitis proportion model used two country level covariates (the proportion of the population living in the meningitis belt and the proportion of the population covered by the MenAfriVac vaccine). The pneumococcal meningitis proportion model used PCV3 vaccine coverage as a covariate and the Hib model used Hib3 vaccine coverage as a covariate. The other meningitis proportion model included Hib3 vaccine coverage, pneumococcal vaccine coverage, and the proportion of people living in the meningitis belt. The four proportion models were scaled to sum to 100% for each location, age-group, sex, and year combination to convert meningitis deaths into meningitis deaths by aetiology [3].
CODEm smooths VR or VA data over time which can result in spikes caused by outbreaks not being represented in the data. For this reason meningococcal meningitis outbreaks were estimated as a fatal discontinuity which means they were added to overall meningitis deaths after they were corrected to fit within remainder of all-cause mortality envelope (CODCorrect). The Global Infectious Disease and Epidemiology Network (GIDEON) and WHO death reports were used as the data sources for epidemic meningococcal meningitis deaths.

Pathogen specific incidence estimates
Overall incidence of acute bacterial meningitis was modelled using DisMod-MR 2.1 informed by incidence data gathered from hospital records, claims data and a systematic review of the literature capturing incidence studies. In total 5,535 site-years of incidence data fed into the model [51]. DisMod-MR 2.1 pools all the available incidence data adjusting for systematic bias associated with the source of the data and its variance from a reference data source which in this case was ICD coded hospital data. Three country level covariates (proportion of the population living in the meningitis belt, coverage of Hib3 vaccine and coverage of MenAfriVac vaccine) were also applied in locations where data was lacking.
Meningitis incidence was then assigned by aetiology using a second set of DisMod-MR 2.1 proportion models. Proportions by aetiology were informed by surveillance data and Proportions of deaths due to each pathogen were informed by a meta-analysis of meningitis case aetiology distribution pre-vaccine era (stratified by region). It was assumed that 88% of meningitis in the pre-vaccine era was caused by pneumococcal, Hib and meningococcal bacteria combined. As the literature on case aetiology distribution was relatively rich and aetiology of mortality distribution relatively poor, poor case aetiology distribution was converted into proportions of deaths using relative pathogen specific case fatality rates (stratified by child mortality setting). Once country and pathogen-specific meningitis deaths prior to vaccine use had been calculated, adjustments were made to account for country specific Hib and pneumococcal vaccine coverage [14].
WHO/MCEE account for children infected with human immunodeficiency virus (HIV) who die from meningitis in HIV/AIDS death estimates [52]. MCEE/JHSPH include deaths from meningitis which occurred in HIV-positive children by applying relative risks for invasive pneumococcal and Hib disease to annual estimates of HIV prevalence [14].
Pathogen specific incidence estimates To calculate meningitis incidence according to pathogen, pathogen-specific deaths according to country were divided by country and pathogen specific meningitis case fatality ratio estimates. In areas with low health seeking behaviour for pneumonia symptoms, as derived from Demographic and Health Surveys and UNICEF's Multiple Indicator Cluster Surveys, a case fatality rate of 90% was assumed.        Figure S1: Quality of underlying cause of death data and modelling methods used to generate death estimates according to model.