Predicting stunting status among under-5 children in Rwanda using neural network model: Evidence from 2020 Rwanda demographic and health survey

Introduction

Stunting remains a significant public health issue worldwide, particularly in low-and middle-income countries. According to the latest estimates by the World Health Organisation (WHO), in 2022, around 148.1 million children under the age of five years about 22.3% of all children in this age group were affected by stunting globally, with the highest burden in low and middle-income countries.¹ The COVID-19 pandemic has exacerbated the situation, since disruptions in food systems and health services are likely leading to an increase in stunting rates.² Therefore, stunting is still a serious public health problem across the world, particularly in Africa.² In 2020, about 58 million children under the age of five were stunted in Africa, accounting for nearly 40% of all stunted children worldwide. The prevalence of stunting in these EAC countries highlights the importance of focused interventions to address the underlying causes of stunting, which include poverty, poor nutrition, and a lack of access to healthcare and sanitary services.³

ANNs are a type of machine learning (ML) technique that has gained prominence in recent years due to its ability to learn and generalise from data, making them ideal for predictive modelling applications.⁴ ANNs are a kind of deep learning, which is a technique that consists of training models with numerous layers of connected nodes to replicate human brain function.⁴ ANNs stand out in significance when contrasted with other ML algorithms for a multitude of compelling reasons. In many circumstances, ANNs can simulate complicated, non-linear interactions between input and output variables, allowing for accurate predictions.⁵ Linear models, such as linear regression, have limitations in capturing nonlinear correlations, but more complex models, such as decision trees or random forests, may overfit the data or be computationally costly, however, ANNs perform better.⁶ ANNs can automatically extract significant features from data, removing the need for manual feature engineering. This can save time and effort while modelling, especially with high-dimensional data.⁷ ANNs are suited for big data applications because they can be scaled to accommodate massive datasets, hence, it is critical to apply them to Rwanda Demographic Health Survey (RDHS).⁷ The RDHS is nationally representative research that collects data on a number of health indicators, including stunting, through house interviews.⁸

In Rwanda, stunting in children under the age of five poses significant health, educational, and economic issues, affecting both individual and national development. Stunted children are more vulnerable to infections and chronic illnesses, increasing morbidity and death rates and putting a strain on the healthcare system.⁹ Stunting-related cognitive deficits result in poor schooling performance and increased absenteeism, lowering adult productivity and earning potential. These impacts reinforce poverty cycles, particularly in rural and disadvantaged regions, aggravating socioeconomic inequities. Recognizing the serious consequences of stunting, the Rwandan government has implemented national strategies to improve maternal and child nutrition as well as strengthen community health initiatives in order to break the intergenerational cycle of malnutrition and promote comprehensive socioeconomic development.¹⁰

Using an ANN to predict stunting offers various advantages, including increased accuracy and the capacity to identify significant predictors of stunting. ANNs can analyse vast volumes of data and detect patterns that typical statistical approaches may miss, allowing for more accurate stunting predictions. Furthermore, ANNs can identify major predictors of stunting, such as poverty, low maternal education, and a lack of access to sanitary facilities, allowing for focused interventions to address the core causes of stunting.¹¹ ANNs have been successfully implemented in predicting disease outcomes, diagnosing medical conditions, and personalizing treatment plans.¹² An ANNs is used in predicting the onset of diabetes and identifying patients at high risk based on a combination of genetic, lifestyle, and environmental factors.^13,14 Another significant application is in imaging analysis, where ANNs have been used to detect and classify abnormalities in medical images, such as tumours in MRI scans, with high accuracy.¹⁵ These applications leverage the pattern recognition capabilities of ANNs to provide insights that might be missed by traditional statistical methods. Furthermore, ANNs have been employed in monitoring patient vitals in real-time, allowing for timely interventions in critical care settings. The adaptability of ANNs to learn from data and improve over time underscores their potential to revolutionize healthcare delivery by enhancing diagnostic accuracy, optimizing treatment strategies, and ultimately improving patient outcomes.¹⁶ A few studies have been conducted in Rwanda using other machine learning technics like logistic regression, Supporting Vector Machine (SVM), Naive Bayes Random Forest (RF), XGBoost gradient model.¹⁷

However, based on the existing knowledge there has been few researches in Rwanda that attempted to utilise ML to predict stunting like the study conducted by Similien et al.¹¹ ANNs have proven to be highly effective in predicting illnesses.^8,18 However, Similien’s publication did not delve into the utilisation of ANNs for this purpose, despite their demonstrated effectiveness in prediction. Recognising this gap, a supplementary study was essential to explore the application of ANNs in predicting stunting in children, using data from the 2020 RDHS. Due to the importance of addressing the root causes of stunting for effective treatments and policymaking, the researcher chose to conduct this study on the application of ANNs in the specific context of stunting in Rwanda, by using the same dataset as the aforementioned publication.¹⁷ The remaining party of this study is organised as follow: Methods, Results, Discussion, Conclusion and recommendation.

Methods

Features importance

Here are 10 steps that we used for feature selection as seen in Figure 1: Step 1 includes importing the dataset and picking the necessary columns for prediction. In this case, the dataset has 23 input features and the ‘stunting’ variable as the target variable. Step 2 the function LabelEncoder is used to encode categorical variables. LabelEncoder is a scikitlearn (RRID:SCR_002577) utility class that encodes category characteristics as numeric values; Step 3 entails separating the data into input (X) and target (Y) variables. The input characteristics are contained in the X variable, while the target variable is contained in the Y variable; Step 4 entails dividing the data into 80% for training (3051 observations) and 20% for testing and validation (763 observations). The ANN model is trained using the training set, and its performance is evaluated using the testing set; Step 5 using Standard scaler which is ideal for data that is approximately normally distributed and is beneficial for models like linear regression and SVM that assume a Gaussian distribution, standardise the input characteristics. The scikit-learn library’s Standard scaler utility class standardises features by eliminating the mean and scaling to unit variance. Step 6: Build an ANN model with two hidden layers and one output layer. The number of neurons in the input layer is equal to the 23 of input features; Step 7 specifically, in our stunting prediction model, we opted for the Adam optimiser and binary cross-entropy loss. The choice of the binary cross-entropy loss function is crucial for binary classification tasks, such as distinguishing between observations of stunting and non-stunting. This loss function quantifies the difference between the predicted and actual outcomes, providing a measure of how well the model is performing in terms of classification accuracy; in Step 8, the model is trained using a dataset comprising a 3814 number of observations. In this case, the training data used for model training involves a determined amount of information. The model is exposed to this dataset over a series of iterations known as epochs. In each epoch, the model refines its weights and biases based on the training data to improve its predictive capabilities. The choice of 100 epochs ensures that the model undergoes a sufficient number of iterations to converge and achieve optimal performance. Additionally, a batch size of 32 is employed, signifying that the model processes 32 observations of data in each epoch before updating its parameters. This batch-wise training approach helps in optimising computational efficiency and contribute to the model’s generalisation ability.

Figure 1. Workflow for calculating and printing Feature Importance using an Artificial Neural Network (ANN).

Step 9 to determine feature importance, the eli5 library’s permutation importance is used. Permutation importance was the model-independent strategy used after training an ANN to predict stunting. It measures the decline in model performance when each feature is randomly shuffled across observations to determine the relevance of particular characteristics. The decline in performance, as measured by measures like as accuracy or AUC-ROC, reflects the significance of a feature. Permutation is performed over numerous samples, and the procedure is iterated to ensure consistency. Features that cause a significant decline in performance are thought to be critical for the ANN’s prediction accuracy in stunting situations. This strategy assists in the identification and prioritisation of critical elements that influence the model’s predictions. For scikit-learn and Keras models, the eli5 module implements permutation importance.²⁹ Step 10 includes printing the features’ importance. The features’ importance is printed in descending order of significance, along with their appropriate weights.

Results

This section presents the results from the ANN model. The total under five children involved in this study were 3814 among them 33.35% were stunted. Figures 2- 4 illustrate key findings, starting with an accuracy of 72% as shown in Figure 2. The ANN model developed in this study strongly predicted stunting among Rwandan children, achieving a 72% accuracy rate and an ROC value of 0.84. These metrics suggest that the model is a valuable tool for identifying children at risk of stunting, as highlighted in Figure 3. Additionally, the feature importance analysis using the ANN model identified negative and positive value characteristics associated with stunting in Rwanda. These insights are crucial for understanding the complex factors contributing to stunting and designing targeted interventions to reduce its incidence.

Figure 2. Model accuracy.

Figure 3. ROC curves of Artificial Neuron Network.

Figure 4. Feature importance using ANN model.

Figure 4 elaborates on the specific factors influencing stunting predictions. Factors negatively correlated with the likelihood of stunting include the mother’s height, the child’s size at birth, the gender of the child, the mother’s education level, the place of residence, reading newspapers, and the mother’s age, all of which tend to indicate a lower risk of stunting. Conversely, factors positively correlated with stunting risk such as the initiation of breastfeeding, the presence of maternal anaemia, marital status, province of residence, occurrences of diarrhoea, altitude, birth weight, and the baby’s age suggest an increased likelihood of stunting, as shown in Figure 4. These findings provide critical insights into the multifaceted nature of stunting and underscore the importance of targeted interventions.

Discussion

The ANN model’s results show considerable promise. The model displays its capacity to correctly categorise cases of stunting and non-stunting within the dataset with a significantly high degree of precision, with an accuracy of 72%. The study on ANN in the prognosis of musculoekeletal diseases show the accuracy ranged between 54.2% to 94.7%. hence the accuracy of 72% is good compared to this study.^11,17

This demonstrates the model’s ability to detect critical patterns and variables associated with stunting in Rwandan children as also shown by Uddin et al.³¹ This paper also show that ANN is very powerful compared to other models.³² Furthermore, the model’s ROC score of 0.84 demonstrates its excellent discriminative capabilities. A higher ROC score indicates better distinction between the stunted and non-stunted child. The algorithm excels at reliably rating stunted children above non-stunted children, as indicated by an ROC score of 0.84, which is crucial in identifying individuals at high risk. These findings highlight the ANN model’s potential as a useful tool for predicting and mitigating stunting in Rwanda.³³ The potential use of ANN is not only used in stunting prediction but also in other healthcare services like an optimum ANN-based breast cancer diagnosis.¹⁵ ANNs have been used to perform a range of data classification and pattern recognition tasks in medical image processing, and they have shown promise as a classification tool for breast cancer. The use of neural networks in image and signal processing grew substantially in the early 1980s.¹⁵ This study highlights the importance of evaluating the ROC curve and AUC-ROC alongside other metrics such as accuracy, precision, and recall to gain a comprehensive understanding of the model’s performance and its practical applicability in predicting stunting. However, it is important to note that this study only used the ROC for assessment.³³

In this study, a multi-layer perceptron (MLP) uses one hidden layer, and it is commonly trained using the backpropagation algorithm, which adjusted the model’s weights by sending errors backward through the network to improve accuracy, as there is a study suggested to use at least one hidden layer.³⁴ Hyperparameters of a model is tuned using methods like grid search or random search. These methods involve testing different combinations of values to find the best-performing set. More advanced techniques, like Bayesian optimization, use probabilistic models to speed up the process. The performance of the model was evaluated on validation data to determine which hyperparameters yield the best results.³⁴

The Adam optimiser is an advanced optimisation method used to train machine learning models, namely neural networks. It combines the benefits of two other prominent optimisers, AdaGrad and RMSProp, by tailoring the learning rate to each parameter separately. Adam maintains track of both the average of the gradients and the squared gradients to adjust the learning rates dynamically during training, which was employed in this work and yield good results.³⁵

Early identification of children at risk of stunting allows treatments to be targeted to those who need them the most, decreasing the burden of stunting on both the individual and society as a whole.³⁶ The use of ANN in stunting prediction offers the potential to improve early detection and intervention tactics. Policymakers and healthcare professionals may prioritise targeted treatments and deploy resources more efficiently if the primary factors leading to stunting are identified.³⁷ The study’s findings suggest that focusing on initiatives to enhance maternal nutrition, promote breastfeeding practices, and improve access to high-quality healthcare services could be of paramount importance in addressing the identified risk factors highlighted in the research as shown in Figure 4. Breastfeeding start appears as an important element, showing that early breastfeeding starting is connected with a decreased chance of stunting. These findings agree with the study conducted with Saberi-Karimian et al.³⁷ Understanding the factors that contribute to childhood stunting gives vital information to policymakers, healthcare professionals, and communities.³⁸

The same as the finding of this study suggested that efforts should be directed at improving maternal nutrition and health, supporting exclusive breastfeeding practices, and guaranteeing access to healthcare services that successfully manage maternal anaemia as well as the prevention and treatment of diarrhoea disorders, the mother’s height is important, as a lower height suggests an increased risk of stunting in children.³⁹ Additionally, customised treatments should be developed for certain provinces with a greater rate of stunting. The awareness among mothers and caregivers about the importance of child proper nutrition, breastfeeding practices, hygiene, and the significance of regular healthcare visits should be increased.⁴⁰ The research recommended that regular health checkups, growth monitoring, and testing to detect potential growth and developmental delays should be improved. Early measures, including nutritional supplements, caregiver counselling, and relevant healthcare interventions, can then be administered.⁴¹ The gender of the child is also associated with the outcome of the stunting of the child as seen in Figure 4. The mothers giving birth to boys should pay close attention to the nutrition of their babies, as different studies revealed boys are more stunted than girls.⁴²

Being at higher altitudes is associated with a high risk of stunting in children in Rwanda as revealed by the study.¹¹ It is crucial to emphasise that altitude is only one of several factors that contribute to stunting, and its influence varies depending on other contextual factors such as economic status, healthcare facilities, and dietary choices.⁴³ To alleviate stunting in high-altitude locations, a comprehensive approach is required, which includes increasing access to healthcare, nutrition, sanitation, and education, as well as addressing the underlying socioeconomic determinants of health,⁴⁴ this study also confirm the findings shown in the different studies. Finally, encourage collaboration among government agencies, healthcare providers, non-governmental organisations, and community-based organisations in order to execute comprehensive and multi-sectoral stunting elimination strategies. This partnership can ensure a comprehensive strategy to address the recognised stunting feature importance by combining the expertise and resources provided by different stakeholders.⁴⁴

However, while the model worked well in this study, it is possible that it may not generalise well to other populations or circumstances. Further study is needed to test the model’s generalisability and to uncover other factors that may improve its performance. Another limitation pertained to the discussion phase, primarily due to the scarcity of existing ANN literature focused on stunting prediction. Consequently, it was challenging to draw meaningful comparisons between this research and prior studies. The contribution of this research is the use of ANN analysis in stunting prediction in Rwanda results in improved identification of significant characteristics, real-time monitoring, targeted interventions, and useful policy decision-making assistance. These contributions increase our understanding of stunting, guide targeted interventions, and may eventually contribute to lowering stunting rates and enhancing children’s well-being in Rwanda.

Conclusions

The feature significance analysis of the ANN model in predicting stunting in Rwanda demonstrates the intricate interaction of numerous factors in affecting child growth and development. Positive value features stress the relevance of breastfeeding habits, mother health, and socioeconomic variables, whereas negative value features emphasise the importance of maternal qualities, education, and environmental factors. Rwanda may adopt potential targeted interventions include nutritional support programs for pregnant and lactating mothers and supplementary feeding for high-risk children, encouraging early initiation and exclusive breastfeeding through community education, enhancing access to regular health check-ups and growth monitoring, providing educational programs for parents on nutrition and hygiene, improving access to clean water and sanitation to reduce infections, implementing special programs in rural areas with limited healthcare access, advocating for policies that support maternal and child health, and developing province-specific interventions based on unique regional needs. Addressing socioeconomic factors through poverty alleviation programs and supporting income-generating activities can also help improve families’ ability to provide proper nutrition, ultimately leading to better child health outcomes. In conclusion, the ANN model developed in this study provides a promising approach to predicting stunting in Rwanda. With further validation and refinement, it has the potential to significantly contribute to efforts aimed at reducing the stunting prevalence and improving child health outcomes in the country.