The role of national nutrition programs on stunting reduction in Rwanda using machine learning classifiers: a retrospective study | BMC Nutrition

Study design

This study focused on a retrospective cohort study. In 2016, the Local Administrative Entities Development Agency (LODA) and Ministry of Local Government (MINALOC) of Rwanda established a Monitoring and Evaluation Information System (MEIS) to monitor and follow up on all households receiving social protection programs across the country. In 2019, LODA, in collaboration with districts, carried out a household survey to address issues of targeting effectiveness through a review of the current ubudehe categories in order to streamline the implementation of social protection programs [33].

This study analysed secondary data from the MEIS system developed by LODA and household survey conducted in 2019 by LODA to examine the benefits of national nutrition programs on stunting reduction among under two years’ children.

Study population and sample size

The study population included all households with ubudehe categories 1 and 2 having at least one child under 2 years of age or pregnant women in 17 Rwandan districts with a high prevalence of stunting and poverty compared to other districts in Rwanda, where the majority of nutrition programs were established. The database of this study contains 92,809 households collected from the integration of two databases, the household survey database and the MEIS database. This household survey was conducted by LODA using administered questionnaire. The questionnaire tool was pre-tested and modified after a pilot study in two districts to ensure the reliability and accuracy of tool, highlight where further training of enumerators is needed. Households with the following criteria were excluded: [1] demographic characteristics of household head( age, sex, education level, wealth category) is missing; [2] lack of information on nutrition and health programs such as nutrition sensitive direct support(NSDS), early childhood development(ECD), antenatal care(ANC) of 4 visits, fortified blended food(FBF), all vaccinations required, Kitchen garden and Toilet facility status; lack of information on nutritional status of child. In total, 8141 including both children and pregnant women met the eligibility criteria and were selected in this study as sample size.

Data collection procedures

In this study, different data sources were employed to compile information on various study variables at the district level. The Rwanda Demographic and Health Surveys were used to determine the prevalence of stunting and the coverage of areas such as vaccination, antenatal care, and child health that may be related to the evolution of the prevalence of stunting [30, 31]. Variables related to social protection programs will be extracted from the MEIS system developed by LODA. Other remaining needed variables will be extracted from a household survey conducted in 2019 by LODA through para-social workers and youth volunteers trained to collect data using questionnaire deployed in their mobile phones.

The terms and conditions for utilizing the household survey data were agreed upon by the researcher, who was prohibited from making disclosure for the household information provided. LODA Board of Directors has approved procedures for using household survey data where the information supplied by household should be treated as strictly confidential and can only be used for research purpose.

Ethical considerations

The study was carried out upon receiving approval from the Local Administrative Entities Development Agency (reference number: NC/NF/290/2022) on July 21, 2022, to access household profiling data and nutrition-sensitive direct support data from system (MEIS). To protect the privacy of study participants, all personally identifiable information was removed during data extraction, and completely anonymous identification numbers were generated.

Statistical analysis

The analysis consisted of descriptive statistics and inferential statistics. Quantitative and qualitative data were analyzed using STATA 13, R, SAS, and Python as statistical tools used to compute coefficients of estimate, survival analysis, and tabulation and evaluation performance of machine learning classifiers. These approaches were selected because they are able to use data from almost any sort of file to create tabular reports and charts, perform descriptive statistics, and carry out sophisticated statistical analysis.

Analysis performed

Kaplan–meier survival curves

The general equation for a KM surviving probability at a stunted time, t_(s), is displayed above. Given survival to at least time t_(s), this formula calculates the probability of surviving past the prior stunted time t_{(s − 1)} by multiplying it with the conditional probability of surviving past time t_(s). 

When substituting for the survival probability, the product of all fractions estimates the conditional probabilities for failure times Ŝ(t_{(s − 1)}) and earlier, the KM formula can also be stated as a product limit [34].

Cox proportional hazard model

To estimate the effects of survival risk-related factors on the stunting status, the cox proportional hazard regression model was utilized. The model is useful in analyzing lifetime data. The continuous random variable (t) in the model represents a person’s lifetime, and the vector of explanatory factors associated with (X) indicates the proportional hazard hypothesis.

According to the cox model formula, the hazard at time (t) is the product of two quantities. The baseline hazard function is h₀ (t), the first of these [34].

Binary logit model specification

The binary logistic regression was used to estimate the coefficients, and odd ratios were used in this analysis to identify the nutrition programs related to stunting reduction in 17 districts of Rwanda after 5 years.The dependent variable is coded as follows: yes = 1 if a child/pregnant woman is not stunted, no = 0 if a child/pregnant woman is stunted.

The child in the selected household is classified as “stunted” or “non-stunted” based on the dichotomous outcome of the user decision, which characterizes the dependent variable (Y). As a result, a household is classified as “non-stunted” when Yi = 1 or as “stunted” when Yi = 0. For such types of dependent variables, either the probit or logit models are appropriate, depending on personal preferences. This model has also been used in an analysis of ethnic minorities’ generational progress in the United Kingdom by examining four labor market outcomes: economic inactivity, unemployment, access to salaried jobs, and self-employment [35]. The binary Logit model was used, and its specifications are as follows:

where logit (p) is the log of the odds \(\left(\frac{p}{1-p}\right)\)

This can also be expressed in terms of probability p, and the model becomes

Z stands for stunting status, which is the outcome dummy variable that indicates whether the child in the selected household is stunted. In this model, ECD attended (X
₁) will be used to estimate the contribution of the ECD program in reducing stunting in Rwanda. The level of poverty in households (X₂) was considered to estimate the effect of poverty on stunting reduction. social protection (X₃) was used to estimate its role in stunting reduction, while access to & use of healthcare (X₄), hygiene & sanitation (X₅), household head age (X₆), household head education (X₇), household head sex (X₈), and kitchen garden (X₉) were used to estimate their contributions to stunting reduction. To capture any measurement error in the stunting reduction, the error term (ε) is appended and left out variables.

when household is not stunted

when household is stunted.

P denotes the probability of not being stunted 1- P is the probability of being stunted.

Other used predicting models

Machine learning algorithms offer effective, model-free solutions to categorize issues. As a result, the performances of various ML algorithms and the statistical classifier were indeed compared. They were selected because most of the variables in the dataset were categorical variables. So, machine learning classifiers helped in evaluating the best method for classification. The ML classifiers that were considered in this study are detailed below.

Decision trees

In decision tree learning, a decision tree is utilized as a predictive model to connect observations about an item to judgments about the target value of the item. A data mining induction technique called the decision tree algorithm repeatedly divides a dataset of records into classes according to whether they are all members of the depth-first greedy approach or the breadth-first approach [36].

Random forest

Random forest is a classification method that focuses on the “rising” of a group of ordered tree classifiers. To classify a new entity, characteristics of this identity are frequently used, employing each classification tree in the forest. The grown trees are built at random, and each tree offers a categorization (or “vote”) for a given class name. The choice is made based on votes cast by most of the forest trees [37].

K-nearest neighbors

Of all machine learning algorithms, the K-Nearest Neighbor Algorithm is the most straightforward. It is founded on the idea that similar samples will typically be found close together. Because they keep all the training samples and wait to build a classifier until a new, unlabeled sample must be classified, instance-based classifiers are also known as lazy learners [38].

Performance criterion

In the study, various evaluation metrics were used to evaluate the prediction models. The criteria are precision, recall score, F1 score, accuracy and AUC score.