School: Los Alamos High
Area of Science: Translational Medical Sciences
Interim: Forecasting Geospatial Acute Malnutrition Prevalence to Optimize Local Procurement of Specialized Nutritious Foods
Lillian Petersen and Garyk Brixi
To see the full version with plots, see https://docs.google.com/document/d/1XZldgjG7CpNL3cQowzVf712PXSuz4ihLHc-gjyjjjfk/edit?usp=sharing
SAM: Severe Acute Malnutrition
MAM: Moderate Acute Malnutrition
SNF: Specialized Nutritious Foods
RUTF: ready-to-use therapeutic food (for SAM treatment)
RUSF: ready-to-use supplemental food (for MAM treatment)
SC+: super cereal plus (for MAM treatment)
Percentages of acute malnutrition continue to be unsettlingly high in developing countries, while coverage of treatment remains unsatisfactory. Treatment and prevention of acute malnutrition typically relies on specialized nutritious foods (SNF) distributed by international agencies. Common SNF is effective but costly with a limited global supply chain.
Insufficient lead time of crop yields and malnutrition prevalence often leads to crucial delays in policy negotiations, planning, producing and trucking food aid. This forces aid organizations to fly in therapeutic food, which significantly increases costs. International aid can also disrupt local markets. In this study, we aim to optimize local supply chains of SNF based on real-time malnutrition estimates to better inform policymakers on how food should be distributed from farmer to factory to child, therefore increasing treatment and sustainability on a lower budget.
Computational Plan and Progress:
The method to optimize the treatment of malnutrition can be broken into three sections: Optimization of locally suitable RUF, predicting malnutrition prevalence, and optimizing local production of RUF, and forecasting of malnutrition.
We developed a linear programming tool to optimize SNF formulae for low cost using locally grown crops. The linear programming tool addresses nutrient requirements conforming to current standards, and optimizes formulae for low cost while allowing users to adjust constraints including nutrients, crop water efficiency, and flavour enhancing ingredients. As a novel approach, it ensures protein quality automatically through protein digestibility corrected amino acid score, balancing proteins with complementary quantities of essential amino acid, to facilitate the removal of costly dairy ingredients.
The tool optimized formulae for international ingredient cost, and then in every sub-Saharan country with available data according to locally grown crops, local prices and crop water footprint, and met all nutrient requirements at a significantly lower ingredient cost and water footprint than current formulae.
The tool’s accuracy was verified against previous recipes, and confirmed through prototyping and laboratory analysis of optimized SNF.
Predicting prevalence of acute malnutrition:
Python computer code was written to predict the geospatial prevalence of malnutrition (wasting, or weight-for-height) throughout sub-saharan Africa. Malnutrition prevalence is partly dependent on low crop yields, so satellite imagery over the growing season was processed to predict crop production in each country 2--4 months before the harvest (see Petersen 2018). These crop yield predictions were then combined with other indicators of political and economic stability, land type, and climate data to predict geospatial malnutrition prevalence. The current training dataset includes (ordered by importance): Percent of females with a secondary school education, distance to coasts, distance to inland coasts, cereal yield, percent of population with access to electricity, population, conflicts, and travel time to the nearest city. This is an initial dataset and will soon include many more indicators. The ground-truth malnutrition data is interpolated from surveys and is published from 2000-2015 in Osgood-Zimmerman 2018.
A random forest regression machine learning algorithm was then fit between a random 80% of the malnutrition data (after being split into boxes to avoid spatial correlations) and the malnutrition covariates from 2000-2015. The machine learning model then predicted the other 20% of the data as a testing set. Soon, we hope to use these same indicators to predict current and future malnutrition prevalence.
Optimizing local production of SNF:
We coded a production network model using the PuLP python library to optimize the supply chain of locally manufactured SNF to meet acute malnutrition treatment demand in sub-Saharan Africa at lowest cost. The model meets the entire existing demand for the combined SAM and MAM treatment in sub-Saharan Africa by selecting the lowest cost local production network while fulfilling demand for treatment. The model selects the number and placement of manufacturing facilities among the 24 countries where local production data was available, the capacity of each factory and its machinery, the amount of RUTF, RUSF, and SC+ produced by each factory, and the destination of produced products. Factory construction and running costs were obtained from a variety of interviews and data sources.
A network with each node representing a country was chosen in order to allow integration with existing national health programmes. This model is easily adjusted according to new data on prevalence, duration of untreated SAM and MAM, duration of treatment, and amount of respective product needed per case. Using the estimated prevalence of acute malnutrition, the production of RUF can be adjusted to meet demand for the following year, and the net cost may be compared to current models.
Results so far:
The optimized SNF are cheaper, more water efficient, and more energy efficient than current formulae, while supporting local production to further sustainability goals. Figure 1 shows current international ingredient cost compared to the optimized international ingredient cost, which is significantly reduced for all SNFs.
Figure 1: International Ingredient Cost Comparison for Specialized Nutritious Foods (see google doc)
Figure 2 shows the water footprint, which was also reduced significantly in the optimized RUTF, RUSF, and SC+ largely due to the removal of milk powder.
Figure 2: International Water Footprint Comparison for Specialized Nutritious Foods (see google doc)
Figure 3 displays the ingredient composition for all of the optimized RUTF recipes based on local ingredient costs, all meeting the micro and macronutrient requirements.
Figure 3: Ingredient Composition Comparison for local optimized RUTF, arranged by cost (see google doc)
Figure 4 shows ingredient cost of the optimized RUTF in between the sub-Saharan African countries and the current international RUTF. All optimized recipes are lower cost than the current international formula.
Figure 4: Ingredient Cost Comparison for local optimized RUTF, arranged by cost (see google doc)
Prototyping in Kenya in collaboration with Valid Nutrition and subsequent professional macronutrient analysis verified the linear programming tool as suitable for development of SNF recipes, as well as verifying taste and consistency of the SC+ recipe. For more information on the linear programming tool, see the recently published paper Brixi 2018.
The malnutrition covariates were fit to malnutrition training data using a random forest regression. The model was trained on a random 80% of boxes across sub-Saharan Africa to avoid spatial correlations (Figure 5a). The model was then asked to predict the other 20% of the data (Figure 5b)
Figure 5: The malnutrition training data in 2015 (a), and the same data with the boxes filled in with predictions (b)
The correlation between the predicted and actual malnutrition prevalence was 0.92 with an average error of 1.73% (Figure 6). We hope to continue to improve these results by adding more indicators.
Figure 6: Predicted and Actual Malnutrition Prevalence (see google doc)
Another future step for this part of the project is to predict malnutrition in real-time using crop and weather forecasts, and using economic and political indicators from the previous year. We also hope to predict stunting (height-for-age) as well as wasting.
Optimizing local production of SNF:
Using locally optimized formulae, estimated costs for transportation, and local production costs, our model met total estimated demand for SAM and MAM treatment with a local production network across sub-Saharan Africa. Given the high transportation cost in Africa, the model recommended 13 factories. The proposed primary supplier of RUTF and RUSF for each country are shown in Figures 7 and 8. The model selected supply zones to meet the demand for both RUTF and RUSF at lowest total cost, taking into account factory placement, capacity, machinery, and product transportation as well as ingredient cost.
Figure 7: Optimized primary national supplier of RUTF (see google doc)
Figure 8: Optimized primary national supplier of RUSF (see google doc)
Except for Cote d’Ivoire and Senegal, the model selected RUTF production in locations with an optimized soybean based recipe, confirming the importance of a low cost source of high protein quality in cost effective RUTF. Production of optimized RUSF included a greater diversity, with both soybean based and peanut based recipes being used. This is due to the ability to cost-effectively meet the lower protein requirement using a wider array of ingredients.
We have several future goals that we hope to add to the supply-chain optimization. Better data on transportation costs, trading corridors, and border crossings will increase accuracy, as well as including shipping as a transportation option (as opposed to only trucking). We also hope to develop a short-term optimization model that uses only current factory locations, from which we could calculate the monetary advantage of the optimized factory locations.
- Bulti, Assaye, et al. “Improving Estimates of the Burden of Severe Acute Malnutrition and Predictions of Caseload for Programs Treating Severe Acute Malnutrition: Experiences from Nigeria.” Archives of Public Health, vol. 75, no. 1, Nov. 2017, p. 66. BioMed Central, doi:10.1186/s13690-017-0234-4.
- Burke, Marshall, and David B. Lobell. “Satellite-Based Assessment of Yield Variation and Its Determinants in Smallholder African Systems.” Proceedings of the National Academy of Sciences, vol. 114, no. 9, 2017, pp. 2189–2194, doi:10.1073/pnas.1616919114.
- Brixi G. Innovative optimization of ready to use food for treatment of acute malnutrition. Matern Child Nutr. 2018;14:e12599. https://doi.org/10.1111/mcn.12599
- Can Yield Gap Analysis Be Used to Inform R&D Prioritisation ScienceDirect. https://www.sciencedirect.com/science/article/pii/S2211912416300621. Accessed 23 Aug. 2018.
- fao.org. “SOFI 2018 - The State of Food Security and Nutrition in the World.” Www.Fao.Org, http://www.fao.org/state-of-food-security-nutrition/en/. Accessed 1 Oct. 2018.
- FOOD & AGRICULTURE ORGANIZATION. STATE OF FOOD SECURITY AND NUTRITION IN THE WORLD 2018. FOOD & AGRICULTURE ORG, 2018.
- Graetz, Nicholas, et al. “Mapping Local Variation in Educational Attainment across Africa.” Nature, vol. 555, Feb. 2018, p. 48.
- Isanaka, Sheila, et al. “Improving Estimates of Numbers of Children With Severe Acute Malnutrition Using Cohort and Survey Data.” American Journal of Epidemiology, vol. 184, no. 12, Dec. 2016, pp. 861–69. academic.oup.com, doi:10.1093/aje/kww129.
- Osgood-Zimmerman, Aaron, et al. “Mapping Child Growth Failure in Africa between 2000 and 2015.” Nature, vol. 555, no. 7694, Mar. 2018, pp. 41–47. www.nature.com, doi:10.1038/nature25760.
- Petersen, L.K. Real-Time Prediction of Crop Yields From MODIS Relative Vegetation Health: A Continent-Wide Analysis of Africa. Remote Sens. 2018, 10, 1726.
- Seal, Andrew. “Mapping Nutrition and Health Data in Conflict-Affected Countries.” The Lancet Global Health, vol. 6, no. 4, Apr. 2018, pp. e365–66. www.thelancet.com, doi:10.1016/S2214-109X(18)30064-0.
- Teravaninthorn, Supee, and Gaël Raballand. Transport Prices and Costs in Africa: A Review of the Main International Corridors. World Bank, 2009.
- The Link between Poverty and Malnutrition : A Household Theoretic Approach. https://ideas.repec.org/p/wbk/wbrwps/536.html#download. Accessed 11 Aug. 2018.
- Weiss, D. J., et al. “A Global Map of Travel Time to Cities to Assess Inequalities in Accessibility in 2015.” Nature, vol. 553, no. 7688, Jan. 2018, pp. 333–36. www.nature.com, doi:10.1038/nature25181.
Sponsoring Teacher: NA
Mail the entire Team