Machine Learning with Neural Networks and Random Forest to Predict Nutritional Risk in Children Under 5 Years of Age

Child malnutrition represents one of the most persistent and alarming challenges worldwide according to the World Health Organization (WHO) [1]. Millions of children face this condition, not only as a consequence of lack of food, but also as a result of multiple social, economic and cultural factors according to their context. This problem, far from being isolated, reflects deep inequalities and requires urgent comprehensive attention [2].

Childhood is a crucial stage in the development of human beings, and adequate nutrition during these first years of life is fundamental to ensure healthy physical, cognitive and emotional growth. However, in many regions of the world, especially in underdeveloped countries, the lack of access to a balanced diet, basic medical care and decent living conditions has turned childhood malnutrition into a silent but devastating threat [3, 4]. From a medical point of view, this condition can have multiple causes, including recurrent infectious diseases such as diarrhea, respiratory infections or intestinal parasites, which hinder the proper absorption of nutrients. In addition, malnutrition is often associated with low birth weight, inadequate breastfeeding practices and deficiencies of essential micronutrients such as iron, zinc and vitamin A, which are essential for strengthening the immune system and cellular development. The risk factors associated with child undernutrition have motivated the search for specific criteria to identify effective alternatives to mitigate this problem [4, 5]. In this context, there is an urgent need to find reliable solutions that contribute to anticipate and address cases of malnutrition, especially in vulnerable populations. This research is justified precisely in the application of innovative technological tools, such as ML, which offer new possibilities to face this challenge from a preventive approach. The use of algorithms such as artificial neural networks and random forest represents a powerful alternative to traditional methods of analysis, since they allow processing large volumes of data and detecting complex patterns that could go unnoticed with conventional approaches [6, 7]. By applying these techniques to the study of child malnutrition in children aged 0-59 months, we seek not only to identify the determinants of its occurrence, but also to generate inputs to guide the formulation of more effective public policies focused on prevention and timely care.

To address this challenge, and in accordance with the proposed objective, we proposed the development of a predictive model using artificial neural networks and the random forest algorithm, implemented through the Python programming language. This tool allowed a rigorous analysis of the data and the generation of accurate statistical visualizations that facilitated the interpretation and understanding of the results obtained. In order to adequately structure each of the stages of the process, the KDD methodology was used, which provided a systematic approach for the selection, cleaning, transformation, mining and evaluation of the data extracted from the selected database [8, 9].

The methodological approach not only guaranteed the quality of the processed data, but also helped to optimize the performance of the algorithms used, increasing the accuracy in the detection of patterns associated with child malnutrition. The fact of using Python because of its wide ecosystem of libraries such as Scikit-learn, Pandas, NumPy and Matplotlib helped to efficiently implement the models, being possible to reproduce the experiments in a controlled and verifiable way [10]. Overall, this combination of statistical techniques, machine learning and exploratory data analysis laid the foundation for a robust technological solution applicable in real-world contexts.

The study aimed to develop a predictive model using neural network algorithms and random forest to identify risk factors associated with child malnutrition in children aged 0 to 59 months, in order to predict its occurrence in vulnerable populations. The achievement of the proposed objective, had as a contribution to make an early prevention of child malnutrition to take corrective measures for a balanced diet suitable for each child.

3.1 KDD methodology

The KDD methodology is an orderly procedure, which consists of a series of phases that allow identifying and extracting patterns and models from the extensive analysis of data collected in an organizational database [34]. Through this ordered way of working, we can process and transform the information into executable knowledge for strategic decisions as shown in Figure 1. The KDD methodology allows iterations between its phases, i.e., to return to previous stages to debug, optimize and refine the data as the analysis develops. This feedback between the different phases of the KDD methodology ensures the validity and relevance of the results obtained, in the sense that the information obtained from its application ends up being valid and useful for data-driven decision making [35, 36].

Figure 1. KDD methodology process

3.1.1 Compilation

For this study, a database that collects information on children with possible indicators of child malnutrition was used in the Kaggle platform. The data include demographic, anthropometric, socioeconomic and access to nutritional intervention programs. Data were obtained from public health records, nutritional surveys, and government databases [37, 38]. Records of children aged 0-59 months were selected, ensuring a representative sample from diverse regions and ethnic groups. Table 1 shows how the variables are classified according to the type of statistical data and the description of each one of them. On the other hand, Table 2 shows the statistical description both as numerical and categorical.

Table 1. Classification of variables

Table 2. Descriptive statistics

In terms of gender distribution, the sample is balanced, with 252 male and 248 female children. Regarding ethnicity, most of the data corresponds to the Xinca population (131), Garifuna (127), Maya (126) and Ladina (116). In addition, it was found that 273 families do not have access to health services, while 227 do. The data show that the average age of the children evaluated is 29.91 months, with an average weight of 11 kg and an average height of 79.51 cm. It has been identified that the monthly family income presents a mean of 272.26 dollars, with a considerable variability reflected in a standard deviation of 130.18 dollars.

The study protocol was reviewed and approved by the Research Ethics Committee of the University of Sciences and Humanities under ACTA CEI No. 018 with approval code 004-2025, complying with national and international regulations for research involving human subjects. Furthermore, the tests performed were based on data extracted from a public database.

3.1.2 Selection, cleaning and processing

It is the value obtained by adding all the data and dividing the result by the total number of observations. It is used to represent a central value in the distribution of the data. As shown in Eq. (1) which is used to calculate the Mean.

$X=\frac{\sum x_i}{n}$             (1)

On the other hand, the dispersion of the data with respect to the mean indicates that the data are highly dispersed, while a low value indicates that the data are more clustered closer to the mean. In that sense, a high standard deviation, for the sample, indicates that the values are highly dispersed, while a low one indicates that the data are closer to the mean using Eq. (2).

$\sigma=\sqrt{\frac{\sum\left(x_i-x\right)^2}{n-1}}$          (2)

Represents the point at which 25% of the data is less than or equal to this value. It is used to analyze the distribution of the data and detect possible outliers in the lower part of the sample. This allows for the presence of outliers if there is a large difference between Q₁ and the median explained in formula (3).

$Q_1=x_{\frac{n+1}{4}}$            (3)

The median is the central value of the sample. It divides the data set into two equal halves. If the number of observations is odd, the median is the middle value; if it is even, it is the average of the two central values. It is interpreted as representing data in distributions where the mean is a good indicator of the central value for both even and odd as indicated by Eqs. (4a) and (4b).

Si n es impar $Q_2=x_{\frac{n+1}{2}}$        (4a)

Si n es par $Q_2=\frac{\frac{x_n+x_n}{2}+1}{2}$          (4b)

Represents the value below which 75% of the data lies. It helps to identify the distribution of the upper part of the data and to detect outliers. The Q₃ value indicates that most of the values are high relative to the mean, to calculate these values the Eq. (5) is used.

$Q_3=\frac{x_{3(n+1)}}{4}$      (5)

Maximum normalization is a method that adjusts the values of a variable so that they fall within a specific range, usually between 0 and 1. This allows different variables to be compared more equitably by establishing a uniform scale for their values, as indicated in Eq. (6).

$X_{\text {norm }}=\frac{X i-X_{\min }}{X_{\max }-X_{\min }}$           (6)

• Data Inspection and Cleaning

The dataset was reviewed to verify variable types (numerical, categorical, and target) and correct any misclassifications, such as numerical values stored as text.

• Handling Missing Data

The percentage of missing values was quantified for each variable. For numerical variables with less than 5% missing data, median imputation was applied; for categorical variables, mode imputation was used. Records with more than 20% missing data across all fields were excluded from the analysis.

• Outlier Detection and Treatment

Outliers in continuous variables were identified using either a z-score threshold of ±3 or the Interquartile Range (IQR) method. For anthropometric measures, extreme values were cross-checked against WHO growth standards. Outliers were treated through winsorization, replacing them with the nearest acceptable limit rather than removing them, in order to preserve sample size.

• Encoding of Categorical Variables

Non-ordinal categorical variables were transformed using one-hot encoding, while ordinal categorical variables were encoded using label encoding.

• Feature Scaling

For the Neural Network model, z-score standardization (mean = 0, standard deviation = 1) was applied to all continuous variables. For the random forest model, the original scales were preserved, as the algorithm is scale-invariant.

• Class Imbalance Handling

The class distribution of the target variable (malnourished / not malnourished) was assessed. In the presence of imbalance greater than 60-40%, the Synthetic Minority Oversampling Technique (SMOTE) was applied, or alternatively, class weights were adjusted (class_weight="balanced" in random forest and weighted loss functions in Neural Networks).

• Data Splitting

The dataset was partitioned into training (60%), validation (20%), and testing (20%) subsets using stratified sampling to maintain the proportion of classes in each split.

The following Figure 2 shows the variability of weight by gender. If the violins have similar shapes, the weights are similarly distributed in both genders; if one is wider in certain areas, it indicates greater concentration of individuals in that weight range. The inner lines represent quartiles, helping to visualize whether a gender has greater dispersion or outliers.

On the other hand, in Figure 3 the dispersion illustrates the relationship between weight and height, where a positive trend is expected: the greater the height, the greater the weight. If the points are well aligned, it indicates a strong relationship; if they are scattered, it suggests greater variability. Differentiation by color makes it possible to analyze whether a gender tends to have a higher weight for the same height, and outliers can indicate possible cases of undernutrition or overweight. Differentiation by color according to gender makes it possible to observe specific patterns between males and females. For example, it can show if one gender has a slightly higher weight compared to the other for the same height, which may be influenced by physiological or contextual differences. In addition, the presence of outliers in the plot points that deviate significantly from the general pattern may point to possible cases of undernutrition, overweight or even errors in data collection. Identifying these outliers is key to performing a more detailed analysis, detecting unusual health conditions and making informed decisions on nutritional or medical interventions.

Figure 2. Weight distribution by gender

Figure 3. Weight distribution by size

The density curves show the distribution of age and weight in the population analyzed. If the curves are similar in shape and position, both values are homogeneously distributed; if one is wider, it indicates greater variability in that aspect as mentioned in Figure 4. The peaks represent the concentration of children in certain ranges, helping to identify growth patterns and possible nutritional deviations.

Figure 4. Age and weight density

3.1.3 Data mining

The implemented model corresponds to a Multilayer Perceptron (MLP) neural network with a feedforward architecture, specifically designed to address the binary classification of child malnutrition. This architecture consists of an initial dense input layer with 64 neurons and a ReLU activation function. It contains the previously normalized variables based on the collected data, thus enabling better convergence during training. Next, an intermediate hidden dense layer with 32 neurons and the same ReLU activation function is incorporated, responsible for capturing complex nonlinear relationships between the child's characteristics and their socioeconomic and health environment (see Table 3). The network concludes with a dense output layer composed of a single neuron with a sigmoid activation function, which returns a continuous probability between 0 and 1, allowing for the determination of whether the child is malnourished or not. Regarding the training parameters, the Adam optimizer was used, known for its efficiency and fast convergence, along with the binary crossentropy loss function, which is ideal for binary classification tasks. Accuracy was adopted as the main evaluation metric to measure the proportion of correct predictions, with a batch size of 8 samples. Additionally, 20% of the dataset was reserved for validation during the training process, allowing for monitoring the model's behavior and preventing overfitting (see Table 4). This configuration demonstrates a robust and well-balanced structure for prediction tasks based on both physical and socioeconomic factors.

Table 3. Multilayer perceptron architecture

Table 4. Training parameters

The Rectified Linear Unit (ReLU) function returns the input value if it is positive, and zero otherwise. It is used to introduce non-linearity into the model and to help learn complex patterns. Eq. (7) refers to the explanation provided above.

$\operatorname{Re} L U_{(x)}=\max (0, x)$            (7)

The sigmoid function transforms any real value into a value between 0 and 1, allowing the output to be interpreted as the probability that a child is malnourished (1) or not (0), as shown in Eq. (8).

$\sigma(x)=\frac{1}{1+e^{-x}}$            (8)

This function penalizes predictions that deviate from the actual value. If the true value $\hat{y}=1$, the model must maximize $\hat{y} ;$ si $y=0$ must be minimized y. t is ideal for binary classification tasks like in this case, as shown in Eq. (9). It measures the percentage of times the model correctly predicted the outcome, as indicated in Eq. (10), for both positive and negative classes. In this case, the calculation is determined based on the area under the ROC curve (AUC), which assesses the true positive rate (TPR) versus the false positive rate (FPR), as shown in Eq. (11). An AUC of 1 indicates perfect classification, whereas an AUC of 0.5 is equivalent to random guessing.

$L(y, \hat{y}\{y\})=-[y \cdot \log (\hat{y})+(1-y) \cdot \log (1-\hat{y})]$           (9)

Accuracy $=\frac{\text { Number of correct predictions }}{\text { Total predictions }}$          (10)

$A U C=\int_{F P R=0}^{F P R=1} T P R(F P R) d F P R$        (11)

• Training Accuracy vs. Validation Accuracy: If validation accuracy starts to stabilize while training accuracy continues to improve, it may indicate overfitting. In this case, the curves appear to be aligned, suggesting that the model generalizes well.
• Training Loss vs. Validation Loss: Similar to accuracy, the training loss decreases over time, which is a sign of the model's learning progress. The validation loss also follows a downward trend, further supporting the idea that the model is not overfitted and is capable of generalizing to new data, as shown in Figure 5.

Figure 5. Training accuracy and loss curves

Previous researchers have considered the effect of the EIR on the nutrition of children in their early childhood through LR models, decision trees, a single-layer neural network, and decision forests. Although this research is very similar to previous ones in terms of the use of the same techniques, it distinguishes itself by employing APP, which has allowed them to reduce selection biases and obtain more robust causal estimates. Additionally, the mediators of women's empowerment and age at marriage were considered, which were not included in previous studies based on CPT [11, 12]. Previous research has primarily focused on estimating infant mortality from demographic and nutritional data, achieving high accuracy rates through the application of random forest, J48, and Naïve Bayes. In this research, the objective is not so much to achieve high prediction accuracy but to understand the causal mechanism between EIR and nutritional deficiencies, identifying sociocultural factors that exacerbate the relationship between EIR and malnutrition [13, 14]. Regarding the model used in this study, the following explains its choice: a deep neural network with a composite structure was developed, in which an initial layer with 64 neurons using the ReLU activation function was implemented, followed by a middle layer with 32 neurons also utilizing the ReLU activation, and finally, an output layer with a single neuron and a sigmoid activation function, designed to predict the probability of childhood malnutrition. The proposed model is able to adequately capture the non-linear relationships between the variables involved in both the socioeconomic and health domains. In contrast to the model proposed by the author [15], who applied Ordinary Least Squares Regression (OLSR), spatial models, and explainable AI algorithms to highlight geographical differences in growth delays, it was determined that while OLSR was useful for identifying linear correlations, the model in this research demonstrated better predictive performance, particularly when incorporating random forest and neural networks. This highlights the potential of adaptive, non-linear artificial intelligence techniques in complex contexts such as childhood malnutrition [16, 17]. In a second study, the authors [18, 19] examined anemia in children under 5 years old, starting with a deep understanding of data mining from a data lake environment. They optimized the preprocessing, cleaning, and variable reduction stages using statistical methods such as ANOVA F-test or chi-square, selecting 10 relevant variables and improving prediction performance [20]. By applying different machine learning algorithms, the Naive Bayes model was determined to be the most efficient for childhood anemia, achieving record (74%), precision (43%), and accuracy (70%) [21, 22]. In this regard, the random forest model developed in this research reached an accuracy of 89%, significantly surpassing these metrics. Additionally, the neural network developed, based on a three-layer architecture, achieved an effectiveness of 87%, highlighting the potential of deep learning and ensemble techniques in identifying childhood malnutrition. Random forest also identified the most determinant variables in the model (weight, height, age, and income), partially aligning with the variables selected by the authors of previous studies, but with greater explanatory power, positioning this model as a more solid alternative for contributing to early public health interventions. Furthermore, the previous studies also proposed a predictive model based on ensemble techniques, conducting the study on a representative sample of children under five years old, applying a standard 70% training and 30% test split, which was the methodology adopted in the present study. Comparing nine different algorithms (random forest, AdaBoost, and XGBoost) also showed the configurations with the highest predictive capacity [23]. In line with these findings, the random forest model applied in this research achieved an accuracy of 89%, standing out as the most effective compared to other individual models. The multilayer neural network used achieved 87%, demonstrating the potential of deep learning as a valid complement. Both studies emphasize the value of combining multiple artificial intelligence techniques to strengthen early diagnosis of childhood malnutrition and improve decision-making in public health. While both studies applied artificial intelligence algorithms to predict childhood malnutrition, this research incorporates a multilayer neural network in addition to classical models such as random forest, SVM, K-Nearest Neighbors, and Logistic Regression, used by other authors [24]. In terms of accuracy, random forest stood out in both studies, but with better performance in this research (89%). The neural network also showed competitive performance (87%), demonstrating that both ensemble models and deep learning are effective tools for early diagnosis of malnutrition. In this research, random forest showed the best performance with an accuracy of 89%, slightly outperforming other models like Logistic Regression (89.88%). The Neural Network achieved 87%, demonstrating its potential, although it did not surpass random forest. Random forest stood out for its ability to handle complex relationships in the data, while the simpler Logistic Regression was less accurate in complex situations. Overall, random forest proved to be the most robust option for predicting children's nutritional status [25, 26]. In the comparative analysis of predictive models for estimating nutritional deficiency in infants, random forest stood out as the most accurate model, achieving an accuracy of 89%. Its main advantage is its ability to handle complex data without the need for highly trained personnel, making it an ideal option in emergency contexts. Logistic Regression, with an accuracy of 89.88%, is simpler but limited by its inability to handle nonlinear interactions. On the other hand, the Neural Network, with a performance of 87%, has great potential for modeling complex relationships, though it requires more data and is harder to interpret. In summary, random forest proved to be the most suitable model for emergency situations, although each model has its strengths and weaknesses depending on the complexity of the context and available resources [27].

The method for extreme and rapid observation of nutritional status (MOERN) uses facial photographs to predict body mass index (BMI) in adults and weight-for-height Z-score (WAZ) in children under five years old. This model achieved an accuracy of 78% in classifying BMI in adults and 60% in a sample of 3167 infants in Kenya [28, 29]. Compared to my research, which used models like random forest and neural networks to predict childhood malnutrition, the use of facial images in MOERN offers an innovative and quick alternative, although its accuracy is still lower than traditional predictive models like random forest (89%) or neural networks (87%) applied in my study. On the other hand, the comparative analysis between VGG16 and SqueezeNet for automatic detection of childhood malnutrition in the cited study [30, 31] presents an approach focused on convolutional neural networks, which aligns with the use of neural networks in my research. However, the Random forest model in my study achieved better performance in terms of accuracy (89%), suggesting that while neural networks are useful, ensemble models like random forest may offer better results in certain contexts, especially in the classification of childhood malnutrition.

The study on the use of VGG16 and SqueezeNet for detecting childhood malnutrition showed that VGG16 achieved an accuracy of 93.73% and a recall of 0.95, making it highly effective for detecting malnutrition. In comparison, SqueezeNet had an accuracy of 90% and a lower recall (0.81). These results highlight the potential of convolutional neural networks for early detection of malnutrition [32, 33].

This study has certain limitations that should be considered. The dataset comes from a specific geographic and socioeconomic context, which may limit the generalizability of the results to other populations. Some potentially relevant variables, such as dietary diversity or healthcare access, were not available and therefore were not included in the analysis. Additionally, although random forest (89% accuracy) and the multilayer neural network (87% accuracy) showed strong performance using socioeconomic and anthropometric data, other approaches such as computer vision models (e.g., VGG16) can outperform them when photographic data is available. However, such models are constrained by the need for high-quality images, which are not always feasible in public health contexts. Despite these considerations, the high accuracy achieved suggests that the proposed models retain strong predictive capacity within the studied context.