WHO lunchtime seminar Mapping child growth failure in Africa between 2000 and 2015 Professor Simon I. Hay March 12, 2018
Outline Local Burden of Disease (LBD) at IHME Child growth failure From global to local o Data o Model o Results Implications and impact Limitations Future directions
Outline Local Burden of Disease (LBD) at IHME Child growth failure From global to local o Data o Model o Results Implications and impact Limitations Future directions
Prevalence of stunting (%) Precision public health: a new approach National (or even subnational) averages can hide important local health variation The use of data to guide interventions that benefit populations more efficiently and increase equity in outcomes National Admin 1 Admin 2 5x5 km
Prevalence of stunting (%) Local Burden of Disease project goals Assemble the world s largest geopositioned dataset on key diseases and risk factors Create high-resolution (5x5 km) maps of prevalence, incidence, or mortality Create compelling and useful interactive data visualization tools to illuminate levels, trends, and disparities over time Disseminate results and encourage uptake by donors, policymakers, and researchers to inform evidence-based decision-making
Location Burden of Disease team 50+ team members o Data Analysts o Data Extraction Analysts o Data Mapping Specialist o Data Services Specialist o o o o o Director Engagement Officer Faculty Fellows PhD Student o o o o Policy Translation Specialist Project Officers Research Coordinator Research Managers o o o Researchers Senior Research Manager Software Engineers
What we are mapping Malaria (P.f. and P.v.) Diarrhea Lower respiratory infections (LRI) Tuberculosis HIV/AIDS Under 5 mortality Educational attainment Child growth and nutrition o o o o Stunting, wasting, underweight Low birth weight Child overweight Exclusive breastfeeding NTD: lymphatic filariasis, onchocerciasis, and schistosomiasis Water and sanitation Vaccine coverage (DTP3, measles, etc.). Ebola and other hemorrhagic fevers Pandemic potential of five emerging zoonotic infectious diseases AMR: 17 bacteria-antibacterial drug combinations
Outline Local Burden of Disease (LBD) at IHME Child growth failure From global to local o Data o Model o Results Implications and impact Limitations Future directions
Child growth failure Stunting Height-for-age z-score <-2 SD Wasting Weight-for-height z-score <-2 SD Underweight Weight-for-age z-score <-2 SD Specific subset of child undernutrition, excluding micronutrient deficiencies Relationship between insufficient height and weight at a given age Described in terms of univariate growth standards by WHO, where agespecific height and weight are compared to healthy reference populations
Future directions Policy relevant analysis WHO Global Targets 2025 to improve young child nutrition Sustainable Development Goal 2.2 to end all forms of malnutrition by 2030, including achievement of the Global Targets 2025
Outline Local Burden of Disease (LBD) at IHME Child growth and nutrition From global to local o Data o Model o Results Implications and impact Limitations Future directions
Geospatial data Point o GPS coordinates (latitude/longitude) o Infinitesimal representation Polygon o Aerial representation (mean over a region) o Typically data matched to shape files Raster o o Data discretized over continuous space, represented by pixel values in a bitmap Covariates and outputs
Data coverage
Sparse data Some areas have robust data coverage, and we can make confident predictions Others have more sparse coverage, so our predictions are less certain
How do we generate predictions in areas with sparse data? 256 geo-located datasets Household surveys A suite of geospatial covariates Satellite imagery and modeled surfaces of relevant environmental and human activity
DATA COVARIATES ENSEMBLE OF MACHINE LEARNING MODELS Maximize the predictive power of the covariates
MODEL-BASED GEOSTATISTICS Borrow strength from observations nearby in space and time, accounting for leftover variation
CALIBRATION TO GBD Leverage validated Global Burden of Disease (GBD) estimates which utilize additional data sources RESULTS Pixel-level estimates with uncertainty intervals, extremely flexible with many use cases
Results 2015, under 5 stunting prevalence National Admin 1 Admin 2 5x5 km
Results 2000, under 5 stunting prevalence
Results 2005, under 5 stunting prevalence
Results 2010, under 5 stunting prevalence
Results 2015, under 5 stunting prevalence
Results 2000-2015, Overlapping populationweighted lowest and highest 10% of pixels and annualised rates of change (AROC) in stunting prevalence
Results Annualized decrease in stunting prevalence from 2000-2015 relative to rates needed during 2015 2025 to meet the WHO GNT Regressing On track Exceeding
Results 2025, Predicted stunting prevalence based on annualised decrease achieved between 2000 and 2015
Results 2015-2025, Acceleration in the annualized decrease in stunting required to meet the WHO GNT by 2025 Met goal by 2015 On track 2x rate of progress needed 4x rate of progress needed
Results 2015, Probability that the Global Nutrition Target for stunting has been achieved at the first administrate subdivision and 5x5 km pixel level
Results 2015, under 5 wasting prevalence National Admin 1 Admin 2 5x5 km
Results 2000, under 5 wasting prevalence
Results 2005, under 5 wasting prevalence
Results 2010, under 5 wasting prevalence
Results 2015, under 5 wasting prevalence
Results 2015, under 5 underweight prevalence National Admin 1 Admin 2 5x5 km
Results 2000, under 5 underweight prevalence
Results 2005, under 5 underweight prevalence
Results 2010, under 5 underweight prevalence
Results 2015, under 5 underweight prevalence
Outline Local Burden of Disease (LBD) at IHME Child growth and nutrition From global to local o Data o Model o Results Implications and impact Limitations Future directions
Implications and impact: publishing
Implications and impact: precision public health
Implications and impact: in the media
Implications and impact: how can decision-makers use the research?
Interactive data visualization tool Explore further at https://vizhub.healthdata.org/lbd/cgf
Outline Local Burden of Disease (LBD) at IHME Child growth and nutrition From global to local o Data o Model o Results Implications and impact Limitations Future directions
Limitations Data coverage and quality o Of 256 data sources, only 127 contain GPS coordinates o Areas of greatest uncertainty correspond to those in need of more/recent information Prediction, not inference o Optimize for prediction, cannot perform correlation inference i.e. relationships between covariates and outcomes Uncertainty propagation o Uncertainty in covariates and population estimates not incorporated
Outline Local Burden of Disease (LBD) at IHME Child growth and nutrition From global to local o Data o Model o Results Implications and impact Limitations Future directions
Future directions: expanding geographic scope Stage 1 + 2 >99% CGF attributable DALYS
Future directions: additional Global Targets 2025
Future directions Further exploring geographic inequalities Recent or forthcoming publications o Educational attainment (Africa) o Diarrhea (Africa) o Water and sanitation (Africa) o Lower respiratory infections (LRI) (Africa) o Under 5 mortality (Global)
Thank you!
Additional slides on methods 5
Model-based geostatistics Model Assume data arises from underlying random process following a known distribution Bayesian hierarchical model Use a generalized linear model framework, which allows us to incorporate covariates (X i ) in our model logit(p i ) = α + X i β + Z i Z GP 0, C 5
Model-based geostatistics Covariates Geospatial team is home to a continually growing spatial covariate repository Both external (e.g. satellite data) and internal (i.e. model outputs) covariates available, in a standardized format 5
Value of interest Model-based geostatistics 1-D example X 1 X 2 X 3 1D index of space 5
Value of interest Model-based geostatistics 1-D example α + X i β + Z i X 1 X 2 X 3 1D index of space 5
Value of interest Model-based geostatistics 1-D example α + X i β + Z i X 1 X 2 X 3 1D index of space 5
Residual Model-based geostatistics 1-D example α + X i β + Z i X 1 X 2 X 3 1D index of space 5
Residual Model-based geostatistics 1-D example α + X i β + Z i X 1 X 2 X 3 1D index of space 5
Value of interest Model-based geostatistics 1-D example α + X i β + Z i X 1 X 2 X 3 1D index of space 6