A deep learning approach for classifying and predicting children's nutritional status in Ethiopia using LSTM-FC neural networks

Abstract

Background

This study employs a LSTM-FC neural networks to address the critical public health issue of child undernutrition in Ethiopia. By employing this method, the study aims classify children's nutritional status and predict transitions between different undernutrition states over time. This analysis is based on longitudinal data extracted from the Young Lives cohort study, which tracked 1,997 Ethiopian children across five survey rounds conducted from 2002 to 2016. This paper applies rigorous data preprocessing, including handling missing values, normalization, and balancing, to ensure optimal model performance. Feature selection was performed using SHapley Additive exPlanations to identify key factors influencing nutritional status predictions. Hyperparameter tuning was thoroughly applied during model training to optimize performance. Furthermore, this paper compares the performance of LSTM-FC with existing baseline models to demonstrate its superiority. We used Python's TensorFlow and Keras libraries on a GPU-equipped system for model training.

Results

LSTM-FC demonstrated superior predictive accuracy and long-term forecasting compared to baseline models for assessing child nutritional status. The classification and prediction performance of the model showed high accuracy rates above 93%, with perfect predictions for Normal (N) and Stunted & Wasted (SW) categories, minimal errors in most other nutritional statuses, and slight over- or underestimations in a few instances. The LSTM-FC model demonstrates strong generalization performance across multiple folds, with high recall and consistent F1-scores, indicating its robustness in predicting nutritional status. We analyzed the prevalence of children’s nutritional status during their transition from late adolescence to early adulthood. The results show a notable decline in normal nutritional status among males, decreasing from 58.3% at age 5 to 33.5% by age 25. At the same time, the risk of severe undernutrition, including conditions of being underweight, stunted, and wasted (USW), increased from 1.3% to 9.4%.

Conclusions

The LSTM-FC model outperforms baseline methods in classifying and predicting Ethiopian children's nutritional statuses. The findings reveal a critical rise in undernutrition, emphasizing the need for urgent public health interventions.

Background

Child undernutrition is a condition characterized by an inadequate intake of essential nutrients, leading to stunted growth, weakened immune systems, and developmental delays [1,2,3]. This condition is commonly assessed using the World Health Organization (WHO) growth standards, which categorize nutritional deficiencies into three main types: underweight (low weight-for-age), stunting (low height-for-age), and wasting (low weight-for-height) [4,5,6,7,8,9,10,11].

Globally, undernutrition significantly impacts young children, with approximately 149 million children under five years old being stunted and 49.5 million wasted, accounting for over 45% of the disease burden in this age group [12, 13]. In Asia, there are 78 million stunted children, including 31 million in India alone. As of 2020, the number of stunted children in Asia was 68 million, closely followed by Africa with 64 million [14]. It was estimated that 21.3% (144 million) of all children under five were stunted, with 36% residing in Sub-Saharan Africa (SSA) and Southern Asia, regions that also account for nearly 90% of underweight children [15,16,17]. Although the prevalence of stunting in SSA decreased from 34.5% in 2012 to 31.1% in 2019, the reduction has not been swift enough to meet global targets [16]. Furthermore, child deaths from undernutrition are nearly double those from diarrhea and pneumonia combined, highlighting the severe impact of this issue [18].

Given this critical global health challenge, requiring ongoing and innovative research methods to fully understand its causes, trends, and future trajectories. In this context, the recurrent neural network (RNN), a recent advancement in deep learning, emerges as a powerful tool for analyzing and predicting patterns in time-series data [19, 20]. Deep learning techniques, including RNNs, have made significant contributions across various domains such as healthcare, finance, meteorology, autonomous driving, natural language processing, and technologies for recognizing images, voices, and handwriting [21,22,23,24,25,26]. A specialized type of RNN, Long Short-Term Memory (LSTM) networks, was introduced by [27] to address the limitations of traditional RNNs, particularly their short memory span and the challenges of vanishing and exploding gradients [28, 29].

The prevalence and related factors of child undernutrition have been studied using various statistical approaches, including regression [30,31,32,33,34,35,36,37], longitudinal models [38, 39], Markov chain [40, 41], Bayesian network [42, 43], and machine learning techniques [44,45,46,47,48,49,50]. However, these traditional statistical methods have limitations in managing long-term dependencies and temporal patterns in sequential data. For instance, Markov chain models assume memorylessness, predicting future states based solely on the current state, which restricts their ability to track and learn from entire data sequences[51, 52]. Similarly, Dynamic Bayesian Networks are limited by their predefined structures and the conditional independence assumptions of each variable, hindering their capacity to model complex, non-linear relationships in data [53, 54]. Traditional machine learning methods, such as decision trees and support vector machines, also struggle to capture temporal dynamics and may inadequately handle sequential data.

In contrast, deep learning models like LSTMs effectively address these challenges through memory cells and gating mechanisms [55,56,57,58,59,60]. LSTMs excel in managing sequential data, capturing complex patterns, and enhancing accuracy in predicting children’s future nutritional status [19, 61,62,63]. Building on this capability, LSTM models are often used to forecast various diseases, such as Parkinson's, bronchopneumonia, dementia, and Influenza-Like Illness [64,65,66,67]. Their success in capturing complex temporal patterns and predicting disease trends highlights their potential. However, the application of LSTM models for child undernutrition remains limited.

Many studies on child undernutrition have employed composite indices [45, 50, 68] or categorized children into groups such as normal, underweight, stunted, and wasted, or by severity as mild, moderate, and severe [44, 46, 69,70,71,72,73,74,75]. For example, machine learning models conducted in countries such as Bangladesh [26, 50, 73, 74], India [47, 76,77,78], Nigeria [74], Gahana [46], and Ethiopia [44, 45, 79] often utilized these approaches. However, these studies did not account for the possibility of concurrent nutritional outcomes, where children might exhibit multiple forms of undernutrition simultaneously. To address this limitation, we employed a more nuanced deep learning classifier approach that considers the probability of children being in multiple states of undernutrition by considering gender difference.

Deep learning classifiers have been applied to early detection of children's nutritional status, as evidenced by studies from Bangladesh [26], Peru [80], India [81,82,83,84], and USA [85]. Most of these studies primarily highlight the effectiveness of deep learning algorithms using images, rather than longitudinal follow-up data. However, despite Ethiopia having one of the highest rates of undernourished children, the application of deep learning algorithms in this context remains relatively underexplored.

Moreover, extensive research has focused on undernutrition among children under five, emphasizing its critical impact during early childhood [4, 5, 32, 36,37,38, 45]. However, studies beyond this stage remain limited, even though early undernutrition often persists, resulting in long-term health and developmental challenges.

As far as we are aware, there is limited evidence regarding the use of deep learning classifier approach in comparison with other baseline methods to assess concurrent nutritional statuses in children from early childhood through adolescence in Ethiopia. Consequently, this proof-of-concept study seeks to explores the potential of an LSTM-FC neural network against models like Auto-Regressive Integrated Moving Average (ARIMA), Random Forest (RF), Recurrent Neural Network (RNN), and Gated Recurrent Unit (GRU), and Bi-LSTM in classifying and predicting undernutrition among children under 15. Our study forecasts the future nutritional status of children over the age of 15, identifying which status is more prevalent and which is less common. As children gain independence and become less influenced by caregivers and families, identifying these trends in nutritional status becomes increasingly important.

Finally, this paper is organized as follows: Sect. 1 introduces the study background, motivation, and objectives. Section 2 details the methodology, including data preprocessing, model development, and evaluation metrics. Section 3 presents the results and analysis. Section 4 discusses the findings in the context of existing literature. Finally, Sect. 5 concludes with key insights, limitations, and recommendations for future research.

Methods

Data source and study design

This study utilized longitudinal data from Ethiopia's Young Lives of Young cohort (YLCS), an international initiative aimed at addressing childhood poverty and health. The cohort includes approximately 1999 children aged 1–15 [86, 87]. The country is highly heterogeneous, with large socioeconomic differences across regional states and between urban and rural areas [88]. The surveys were conducted at 20 sentinels across five Ethiopian regions: Amhara, Oromia, Southern Nations, Nationalities, and Peoples (SNNP), Tigray, and Addis Ababa City Administration (CA). Data collection occurred in five distinct waves—2002, 2006, 2009, 2013, and 2016—each spaced at approximately equal intervals throughout the study period [89]. The original dataset consists of 9,995 observations, representing 1,999 children, with each child having 5 visits, including a zero placeholder for any missed visits. The training data initially includes 8,556 observations, but after applying SMOTE for data augmentation, the training set expanded to 30,208 observations, while the testing data remains at 1,440 observations.

Intervention programs were available across the study regions, targeting specific vulnerable populations. The Productive Safety Net Program (PSNP) operated in 14 sentinel sites across four regions (excluding Addis Ababa City Administration), focusing on pro-poor households [90]. Meanwhile, the Emergency Aid Program (EAP) and Health Extension Program (HEP) were available in all five regions, addressing the needs of disadvantaged socioeconomic groups by offering antenatal care, childhood disease management, and micronutrient supplementation [91].

To identify beneficiaries, the study conducted interviews with randomly selected households to assess their participation in the PSNP, EAP, and/or HEP within the past 12 months. Households were categorized as beneficiaries or non-beneficiaries for PSNP and EAP beginning in 2009 (third wave), while HEP categorization started in 2013 (fourth wave). These intervention programs were later consolidated into a single variable, “program participation status,” encompassing eight categories (C, P, E, H, PE, PH, EH, and PEH), as detailed in Table 1 at [41].

Table 1 Performance metrics of LSTM-FC model across different time-based data splits

Variables in the study

Outcome variable

The YLCS utilized WHO standards to standardize Z scores for each observation, defining children's anthropometric status based on categories such as underweight, stunted, or wasted. After cleaning and making necessary adjustments to the dataset, we classified each child's anthropometric condition at each time point: U for underweight only, S for stunted only, W for wasted only, US for underweight and stunted, UW for underweight and wasted, SW for stunted and wasted, and USW for all three conditions. Furthermore, the condition is classified as Normal (N) if the child is not underweight, stunted, or wasted, indicating their measurements are within the expected range for their age and sex [92]. The classification percentage of children's nutritional status over time are presented in Table S2.

Features

The list of features (independent variables) included in this study were extracted from the dataset and have been extensively discussed in the literature [93,94,95,96,97]. These variables serve as an input feature ($x$) to the LSTM-FC neural network, which processes them internally through its gates to model temporal dependencies and make predictions. A list of these independent variables, along with their distribution, bivariate analysis and corresponding visualizations, is provided in the supplementary material.

Data preprocessing pipeline

The research workflow starting with data preprocessing, which involves handling missing values and identifying outliers to ensure analysis reliability before proceeding to feature selection (Fig. 1). To handle missing values, we applied mean imputation for continuous variables and mode imputation for categorical ones, repeating the process until all variables were fully complete [98]. Furthermore, string data was converted to numeric values using Python's widely-used ‘Numpy’ library. The categorical data was converted using one-hot encoding to make it compatible with deep learning algorithms in the LSTM-FC model. Similarly, as detailed in the supplementary material, Min–Max normalization was applied to scale the features. This process enhanced data consistency, leading to improved model performance and increased prediction accuracy.

Fig. 1

Workflow Diagram of the LSTM-FC Model for Predicting Child Undernutrition in Ethiopia (2002—2016) [26, 99]

Furthermore, we employed the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance in the dataset, as outlined in the supplementary material. SMOTE generates synthetic samples for minority classes, promoting a more balanced class distribution. This technique enhances model performance and prediction accuracy, particularly for minority classes, by reducing bias and improving the robustness of the results. Initially, the training dataset consisted of 7,582 samples, with 3,706 instances representing the majority class (normal) and only 6 instances for the minority class (stunting and wasting concurrently). After applying SMOTE, the training dataset was expanded to 29,648 samples (Fig. S3). This augmentation not only balanced the class distribution but also improved the model's ability to learn from the minority class.

For feature selection, we employed SHAP (Shapley Additive Explanations), a method that helps to identify the most impactful features by providing insights into the contribution of each feature to the model's predictions. This approach enhances interpretability and allows for a more informed selection of features, ensuring that the model focuses on the most relevant variables. Features with higher absolute SHAP values are considered more influential in the model's decision-making process. This approach aligns with established practices for feature selection in LSTM models, enhancing the reliability and robustness of our results [85, 100, 101]. As shown in Fig. 2, the mean absolute SHAP values for features above the median are presented. Due to publication length constraints, SHAP values for all variables are provided in supplementary material.

Fig. 2

Mean Absolute SHAP Value (average impact on model output) for features above the median). Note: Mean absolute SHAP values for all features are provided in the supplementary material due to publication length constraints

In this study, a time-based data splitting strategy was employed to account for the longitudinal nature of the dataset and ensure temporal consistency between training and testing, and align with the requirements of the LSTM-FC model. A recommended approach for longitudinal data is to train the model using earlier time points and test it on the latest data, as this reflects real-world forecasting scenarios where historical data is used to predict future outcomes. This approach ensures the model learns temporal dynamics over time and is evaluated on its ability to generalize to unseen, future data.

Architecture of LSTM-FC neural network

In this study, we used an LSTM-FC model to analyze sequences of health and socio-demographic data over time to predict trends in child undernutrition. As illustrated in Fig. 2, LSTM-FC networks employ several key gates to manage the flow of information.

• a Forget Gate (${f}_{t}$): The forget gate determines which parts of the previous cell state (${c}_{t-1}$) to retain or discard, filtering out unnecessary historical information about a child's nutritional status for predicting future outcomes. The forget gate is computed using the formula [102]:

  $${\text{f}}_{\text{t}}=\upsigma \left({\mathbf{W}}_{\text{f}}\cdot \left[{\text{h}}_{\text{t}-1},{\text{x}}_{\text{t}}\right]+{\text{b}}_{\text{f}}\right)$$

  (1)

  where $\sigma$ is the sigmoid activation function, which compresses the output between 0 and 1; ${{\varvec{W}}}_{f}$ is the weight matrix for the forget gate; ${b}_{f}$ is the bias term; ${h}_{t-1}$​ represents the hidden state from the previous time step; and ${x}_{t}$ is the current input.

• b Input Gate (${i}_{t}$): This gate determines which new information from the current input or features ${(x}_{t})$, and the previous hidden state ${(h}_{t-1})$ should be added to the cell state. It manages the updating process by regulating the incorporation of new features into the cell state. The input gate can be calculated using [103]:

  $${\text{i}}_{\text{t}}=\upsigma \left({\mathbf{W}}_{\text{i}}\cdot \left[{\text{h}}_{\text{t}-1},{\text{x}}_{\text{t}}\right]+{\text{b}}_{\text{i}}\right)$$

  (2)

  where ${{\varvec{W}}}_{i}$ is the weight matrix for the input gate, and ${b}_{i}$ is the corresponding bias term.

• c Candidate Cell State (${\widetilde{c}}_{t}$): This component generates a vector of potential new values that could be integrated into the cell state, representing potential changes in the child's nutritional status based on new data inputs. Combined with the input gate, it contributes to updating the cell state. The candidate cell state is given by [103]:

  $${\widetilde{\text{c}}}_{\text{t}}=\text{tanh}\left({\mathbf{W}}_{\text{c}}\cdot \left[{\text{h}}_{\text{t}-1},{\text{x}}_{\text{t}}\right]+{\text{b}}_{\text{c}}\right)$$

  (3)

  where ${{\varvec{W}}}_{c}$ is the weight matrix for the candidate cell state; ${b}_{c}$ is the bias term; and $\text{tanh}$ is the hyperbolic tangent activation function, which scales values between −1 and 1.

• d Cell State Update (${c}_{t}$): The cell state ${(c}_{t-1})$ is updated by merging the previous cell state ${(c}_{t-1})$, forget gate output ${(f}_{t})$, and the candidate cell state${(\widetilde{c}}_{t})$ . This update reflects the current understanding of the child's nutritional status, combining past trends and new information. It serves as a memory that helps retain information over long sequences. The update is performed as [102]:

  $${c}_{t}={f}_{t}\odot {c}_{t-1}+{i}_{t}\odot {\widetilde{c}}_{t}$$

  (4)

  Here, ⊙ denotes element-wise multiplication, which means that two vectors or matrices of the same dimensions are multiplied together element by element.

• e Output Gate (${o}_{t}$): This gate determines which elements of the updated nutritional state should be used to make predictions about the child's nutritional status in the next time step. It determines what information from the cell state will be passed on to the next layer or time step. It can be computed as [103]:

  $$O_t\;=\upsigma\;(W_o\;\cdot\;\lbrack h_{t\;-\;1,}\;x_t\rbrack\;\;+\;b_o)$$

  (5)

  where ${{\varvec{W}}}_{o}$ is the weight matrix for the output gate, and ${b}_{o}$ is the bias term.

• f Hidden State ($h_t$): The final output of the LSTM cell is produced, which could represent the predicted likelihood of undernutrition or the expected nutritional status of the child at the next time step, based on the current cell state $\left(c_t\right)$ and the output gate $\left(O_t\right)$. The hidden state is calculated as [102]:

  $${\text{h}}_{\text{t}}={\text{o}}_{\text{t}}\odot \text{tanh}\left({\text{c}}_{\text{t}}\right)$$

  (6)

  where ${h}_{t}$ represents the hidden state at time $\text{t}$, $\text{tanh}\left({c}_{t}\right)$ is the hyperbolic tangent of the cell state at time $\text{t}$, and ${o}_{t}$ is the output gate at time $\text{t}$, and the symbol ⊙ denotes element-wise multiplication, combining the output gate with the tanh of the cell state to produce the hidden state.

As shown in Formula 7, the sigmoid function ($\sigma$) is used as the activation function for the gates ${i}_{t}, {h}_{t}$, and ${o}_{t}$ to ensure that their values stay within the range of 0 to 1 [65].

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

(7)

The hyperbolic tangent function (tanh) function, as shown in Formula 8, is used as an activation function for the cell state update to normalizes the input to a range between −1 and 1, effectively centering the data and addressing vanishing gradient issues in neural networks. Its output is centered around 0, and with a larger gradient near 0, this helps the model to converge more quickly [65].

$$\text{tanh}(x)=\frac{{e}^{x}-{e}^{-x}}{{e}^{x}+{e}^{-x}}$$

(8)

We used the Rectified Linear Unit (ReLU) activation function for the fully connected layer, shown in Formula 9 to prevent the vanishing gradient problem, as it maintains a gradient of 1 for positive inputs, which helps in accelerating training and improving model performance [65].

$$\text{ReLU}(x)=max(0,x)$$

(9)

The Softmax function (Formula 10) is used in the output layer for multi-class classification. It converts raw scores into probabilities for each of the 8 nutritional statuses, enabling the model to predict the likelihood of each category effectively.

$$\text{softmax}\left({x}_{i}\right)=\frac{{e}^{{x}_{i}}}{\sum_{j} {e}^{{x}_{j}}})$$

(10)

Algorithmic explanation for LSTM-FC network architecture

The architecture of our LSTM-FC model, illustrated in Fig. S2, features two hidden layers with 256 and 128 neurons, and a fully connected layer with 128 neurons. The detailed algorithmic explanation is provided as follows:

Algorithm 1. LSTM-FC Neural Network Architecture

Model training strategy

During training, the dataset was split into training and testing sets, with the model trained over a maximum of 500 epochs. The model's dimensions were defined by batch size, time step, and feature count. The Adam optimizer was used with an initial learning rate of 0.001, and learning rate scheduling was applied to adjust the rate as training progresses [104]. To prevent overfitting, regularization techniques specifically dropout with rates ranging from 0.2 to 0.5 were employed. The training process is monitored using validation loss and accuracy, with early stopping implemented to halt training if there is no improvement after a specified number of epochs. Finally, the model was evaluated on the test set to assess its generalization ability, and hyperparameter tuning was performed to optimize its performance.

Hyperparameter tuning

Hyperparameters are key to model performance. Throughout the training phase, we explored a range of hyperparameters such as learning rate, number of layers, number of neurons, dropout rate, batch size, number of epochs, optimizer, and activation function. The learning rate controls weight update speed, batch size affects training efficiency, and the number of layers and units per layer shape the model’s complexity. Dropout rate prevents overfitting, while epochs determine how often the dataset is fully processed. Proper tuning of these hyperparameters is essential for achieving optimal results. To determine the optimal number of hidden layers, we train models with varying configurations, starting with one layer and increasing complexity incrementally. We assess performance on a validation set and continue adding layers until gains are minimal. Similarly, for the number of neurons, we typically use powers of 2 (e.g., 64, 128, 256) for simplicity and ease of comparison.

Evaluating the model's ability to generalize

The generalization capability of the trained LSTM, which refers to its ability to accurately classify new and previously unseen data is crucial. To assess this, the classification error E^class is evaluated on the test set, which contains data that was not used during the training process. This allows for an unbiased evaluation of how well the LSTM model generalizes to new data. The classification error E^class is typically calculated by comparing the predicted labels with the true labels of the test set. A lower classification error indicates better generalization performance, as it suggests the model has learned patterns that can be applied to unseen instances. Furthermore, we employed eightfold cross-validation, where the entire dataset was divided into eight subsets (folds). In each iteration, one-fold was used as the testing set while the remaining seven folds were used for training the model. This process was repeated for all folds, ensuring that every data point was used for both training and testing, providing a robust estimate of the model's performance across different data splits.

Model performance metrics

In this paper, we used several performance metrics to evaluate the effectiveness of our model. These metrics include accuracy, sensitivity (recall), specificity (precision), F₁ score, and confusion matrix, each providing a distinct perspective on the model's classification performance. The detailed definitions and explanations of these performance metrics, along with their respective formulas, are provided in the supplementary material. We evaluate the performance of the LSTM-FC model against several baseline methods: (I) the ARIMA, (II) RF, (III) RNN, (IV) the GRU, and (V) Bidirectional LSTM model, with detailed descriptions provided in the supplementary material.

Computational environment

In this study, we used Python's TensorFlow and Keras libraries to leverage their powerful neural network capabilities for training and testing our LSTM-FC model [105] on a system equipped with an Intel I7-9300H CPU, NVIDIA GTX 1650 GPU, and 32 GB RAM.

Results

The nutritional status of children in Ethiopia, as detailed in Fig. 3, shows significant fluctuations over time. During infancy (< 2 years), the proportion of children with a normal nutritional status peak, with 19.37% of females and 18.32% of males in this category. However, by early adolescence (12–14 years), this figure decreases to 15.12% for females and 16.63% for males. In contrast, as children grow older, severe undernutrition (USW), which includes underweight, stunted, and wasted conditions concurrently, increases. For males, this issue rises sharply during middle adolescence (14–18 years), affecting 47.59%. In contrast, females experience the highest rate of severe undernutrition earlier, with 52.34% affected during early adolescence (12–14 years). The detailed values can be found in the supplementary material, Table S4. Figure 4.

Fig. 3

Computational diagram of a LSTM-FC Cell Architecture [102, 106]

Fig. 4

Distribution of child nutritional status by their developmental stage and gender in Ethiopia (2002 – 2016). The values in parentheses give the percentage (%)

Comparison of metrics across different dataset splits

The dataset, collected across five time points was divided into various training and testing splits to evaluate model performance under different scenarios. Among the evaluated splits, training on the comprehensive dataset from 2002, 2006, 2009, and 2013 and testing on 2016 yielded the best performance, with high sensitivity (0.85), specificity (0.84), F1-score (0.84), AUC of 0.90 and an accuracy of 0.83 (Table 1).

Performance evaluation of LSTM-FC model against various baseline methods

In this section, we assess the predictive accuracy and long-term forecasting capability of the LSTM-FC model for evaluating child nutritional status. The model's performance is compared across various developmental stages, including infancy, early childhood, middle childhood, early adolescence, and middle adolescence, against several baseline models (Fig. 5).

Fig. 5

The prediction results of the LSTM-FC model and other baseline prediction horizons on child nutritional dataset in Ethiopia

Our findings reveal that neural network-based approaches, particularly the LSTM-FC and Bi-LSTM models, excel in capturing temporal features and consistently outperform traditional baseline models such as RF, ARIMA, RNN, and GRU in terms of prediction accuracy and precision. For example, in the 2016 nutrition status forecasting task, the LSTM-FC and Bi-LSTM models achieve approximately 36% and 34% higher accuracy than the ARIMA model, with precision improvements of about 25% and 23%, respectively (Fig. 5). Additionally, these models exhibit substantial enhancements across key performance metrics, including accuracy, precision, recall, and F1-Score, when compared to the baseline methods.

The LSTM-FC model demonstrates superior performance across various evaluation metrics and prediction horizons, as illustrated in Fig. 5. It consistently achieves the highest metrics, regardless of the prediction horizon, and shows minimal fluctuations in results. This stability indicates that the model is robust and unaffected by horizon variations, making it suitable for both short-term and long-term predictions. Detailed performance values for each model are provided in Table S6 of the supplementary material for reference. Overall, the LSTM-FC model consistently outperformed ARIMA, RF, RNN, GRU, and Bi-LSTM-FC across all time periods, achieving the highest overall accuracy (93.4%), precision (91.4%), recall (89.8%), and F1-score (90.2%), demonstrating its superior capability in evaluating child nutritional status.

Classification of child nutritional status across gender: confusion matrix analysis

To further analyze the model's performance, we present the confusion matrix for each nutritional status category across gender (Fig. 6). This comparison highlights the model's performance in predicting nutritional status combinations, revealing potential differences in classification accuracy between male and female children. Each cell represents the count of instances predicted for a specific category. The diagonal values indicate correct classifications, while the off-diagonal values represent misclassifications between different nutritional status categories.

Fig. 6

Confusion matrix for child nutritional status classification model across gender

The confusion matrix in Fig. 6 visualizes the performance of the classification model, comparing true versus predicted nutritional status categories across gender. The model performs exceptionally well, with the normal category (N) and SW showing zero misclassifications, as indicated by the absence of off-diagonal values in that row for both males and females. The diagonal values for other categories are also significantly higher than the off-diagonal values for both genders, further showcasing the LSTM-FC model's accuracy in classification. Overall, the model demonstrates strong predictive accuracy for most nutritional status categories across gender.

The model was trained to predict the nutritional status of children over 500 epochs, with a maximum of 4000 iterations, utilizing 8 iterations per epoch on a system equipped with a GPU. To gain a deeper understanding of the model's performance, we analyze the prediction errors for each nutritional status category across gender, as shown in Fig. 7. Each point in the plot represents the prediction error for a specific category, where the error is the difference between the true and predicted nutritional status. The distribution of errors highlights the model's performance, with certain categories showing tighter error ranges compared to others.

Fig. 7

Performance of nutritional status prediction across gender

The N (normal) and SW (stunted and wasted concurrently) categories exhibit a mean error close to zero, indicating accurate predictions with minimal bias, consistent with the findings in Fig. 6. In contrast, most categories show negative errors in the range of 15 to −20, suggesting slight underestimation of nutritional status in those categories. With the exception of the U (underweight) and UW (underweight and wasted concurrently) categories, all other nutritional status categories exhibit negative prediction errors for both genders. Specifically, in females, the US (underweight and stunted simultaneously) category demonstrates a notably large prediction error, whereas in males, the S (stunted) category exhibits the largest prediction error.

Prevalence prediction percentage

The plot presents the nutritional status of children across different age groups, with the percentage of children affected by various conditions. For instance, the normal (N) category shows a gradual decline in the percentage of children with healthy nutritional status from 46.2% at age 1 to 32.7% at age 25. The underweight (U) category fluctuates slightly, peaking at 5.9% at age 8 and declining to 2% at age 1 and 15% at age 25. The stunted (S) condition peaks at 50.8% for children between the ages of 8 and 15, gradually declining to 3.8% by age 25. The wasted (W) condition starts at 3.2%, peaks at 7.2% at age 15, and stabilizes around 6.9% at age 25. In the second subplot, the underweight and stunted (US) category fluctuates, with a peak of 18.6% at age 1 and a decline after the early years. The underweight and wasted (UW) category starts at 6.1% at age 1, increases to 19.4% at age 15, and then decreases. The stunted and wasted (SW) category remains low, with percentages ranging from 0 to 0.7%. The underweight, stunted, and wasted (USW) category starts at 7.4% at age 1, peaks at 14.4% at age 12, and ends at 9.2% by age 25 (Fig. 8).

Fig. 8

Prevalence prediction of Ethiopian children nutritional status from age 1 to 25

Evaluation of model’s generalizability

The progression of E^class, assessed on the training dataset, is shown in Fig. 9(a). In runs incorporating both classification and prediction task, the error decreases significantly more rapidly, resulting in a much smaller classification error after training. Figure 9(b) presents the success rate of finding a solution with E^class ≤ 5%, as a function of the maximum number of epochs. Success rate is measured on both the training and test datasets. When incorporating the additional source of information, 96% of the runs achieve an Eclass on the training set reaches the 5%, and 94% of the runs meet this threshold on the test set. These results indicate a substantial improvement in model generalizability, as the rate is almost higher than when the training process relies solely on the classification task.

Fig. 9

Generalizability of LSTM-FC Model

To further evaluate the generalizability of the model, eightfold cross-validation was performed, and the evaluation results for each fold are reported in Table 2. This approach ensures that the model's performance is assessed on diverse subsets of the data, providing a comprehensive understanding of its generalization ability. The results show varying performance across the folds, with Fold 5 consistently achieving the highest scores in Precision, Recall, Accuracy, and F1-Score, indicating its superior generalization capability compared to the other folds. These findings are further supported by the training results and visualized in Fig. 9, demonstrating the model's effectiveness in both classification and prediction tasks.

Table 2 Generalization performance scores of the LSTM-FC model

Discussion

Previous research indicates that Ethiopia ranks among the countries with the highest number of undernourished children globally and in Sub-Saharan Africa [107,108,109,110,111,112]. Although there has been a decline in under-five undernutrition in Ethiopia, it has not met Millennium Development Goals expectations [112]. Undernutrition in the first 1000 days (from birth to 23 months) often persists until age five and beyond, leading to long-term physical, psychological, cultural, and socioeconomic impacts [113]. Furthermore, the prevalence of undernutrition into later developmental stages is under-researched globally, including in Ethiopia [114, 115].

In comparing the performance of LSTM-FC across various predictive horizons, our study demonstrates the clear superiority of the LSTM-FC model over other baseline models, including ARIMA, RF, RNN, GRU, and Bi-LSTM-FC, at all-time points. The LSTM-FC consistently achieves higher metrics, emphasizing its effectiveness in capturing temporal dependencies and complex sequential patterns in longitudinal data. These findings align with existing literature [116, 117], which highlights the strength of LSTM models in retaining long-term dependencies. While models like GRU and Bi-LSTM-FC also perform well, their metrics fall slightly short of LSTM-FC, particularly in precision and recall, indicating a less balanced trade-off between false positives and false negatives.

In a study conducted by Sima Siami-Namini et al. [118], Bi-LSTM demonstrated superior predictive performance compared to ARIMA and LSTM models; however, in our study, LSTM-FC outperformed Bi-LSTM with slightly better prediction, highlighting its enhanced ability to model temporal dependencies in longitudinal data. Similarly, in the study by Pirani et al. [119], a comparative analysis of ARIMA, GRU, LSTM, and Bi-LSTM was conducted on financial time series forecasting. The authors demonstrated that LSTM-based models, particularly Bi-LSTM, outperformed traditional methods like ARIMA. While Bi-LSTM provides enhanced training with its bidirectional structure, the study highlights the importance of choosing the right algorithm for minimizing errors and achieving superior predictions. In contrast, our study finds that LSTM-FC slightly outperforms Bi-LSTM, further emphasizing the effectiveness of LSTM-based architectures in time-series prediction tasks.

Optimizing the number of hidden layers and neurons is crucial for enhancing our LSTM-FC model's performance and efficiency. More layer’s help capture complex patterns, and more neurons allow for processing extensive information [29]. The right balance prevents overfitting, reduces computational demands, and shortens training times [120]. Through extensive testing, we identified the optimal layer and neuron configuration for our dataset. As a result, the LSTM-FC model's architecture, featuring two hidden layers followed by a fully connected layer, strikes an optimal balance between complexity and computational efficiency. This design, consistent with findings that two to three hidden layers are typically sufficient for most applications [26, 65, 121,122,123].

The research conducted by Amin and Novitasari [124] used an LSTM model to identify stunting in hospital data based on gender, age, and height. This approach leverages the LSTM's ability to handle sequential data, aiding in the early detection and diagnosis of stunting. However, in our study, combining LSTM with fully connected (FC) layers outperformed the standard LSTM and bidirectional models, yielding better results for predicting child nutritional status. Addressing overfitting challenges, our model uses a single FC layer with a dropout ratio of 0.3, differing from Zhao et al.'s approach of using two FC layers with a dropout of 0.5 and 0.3 [65].

In conclusion, the use of LSTM-FC in healthcare is still relatively limited. Notably, Zhao et al. [65] employed a deep neural network combining LSTM and FC layers to predict bronchopneumonia in pediatric patients. Additionally, LSTMs have been effectively used in other applications, including the identification of surgical site infections (SSIs) from electronic health records (EHRs) [125]. Similarly, our research demonstrates the LSTM-FC model's superior performance in handling long-term dependencies in time-series data, highlighting its potential in the dynamic field of nutritional epidemiology. However, the implementation of the LSTM-FC model does require significant computational resources, which may pose challenges in low-resource settings.

This paper was aimed to predict the prevalence of nutritional status of children beyond age 15 using the LSTM-FC model. It analyzes data from the Young Lives study in Ethiopia, which tracked children's nutritional status from age 1 in 2002 to age 15 in 2016 through five survey rounds, with data collected approximately every 3.5 years. This study serves as a benchmark, demonstrating the effective implementation of LSTM-FC algorithms to predict nutritional status as children mature.

Conclusion

Research on the early prediction of children's nutritional status is crucial, as undernutrition remains a critical global health issue, particularly affecting developing countries such as Ethiopia. To the best of our knowledge, this study is the first to investigate the use of an LSTM-FC neural network model for classifying and predicting nutritional status of Ethiopian children, showcasing its superior performance over baseline models. The LSTM-FC model achieves high accuracy in classification, with most predictions correct and only a few misclassifications observed in both male and female categories. It performs particularly well in identifying normal and stunted-wasted (SW) nutritional statuses. Furthermore, the prediction error is also minimal, with a slight underestimation observed, specifically for the stunted category in males and the US category in females. Despite this, the overall classification accuracy remains strong across both genders. Likewise, the evaluation of the model’s generalizability highlights its robust performance across various nutritional categories, with Fold 5 yielding the best results, while minimal prediction errors. Finally, our findings reveal a concerning decline in the healthy nutritional status of children, with a notable increase in the prevalence of multiple undernutrition conditions by 2027. These results underscore the urgent need for targeted public health interventions and policies aimed at combating undernutrition among Ethiopian children. Given Ethiopia's challenges with healthcare facilities, food insecurity, and displacement due to political instability and natural disasters, it is imperative for the government, private sectors, and policymakers to take decisive action.

Strengths, limitations, and future work

We compare the performance of the LSTM-FC model with existing baseline methods, demonstrating its superiority in terms of classification accuracy, generalizability, and minimal prediction error across various nutritional categories. Additionally, the study allocates sufficient time for extensive hyperparameter tuning, ensuring the model achieves optimal performance with minimal error. Similarly, the paper underwent rigorous data preprocessing steps to ensure the quality and reliability of the dataset. Finally, the paper considers the concurrent nutritional conditions of children, accounting for multiple undernutrition states simultaneously, which provides a more comprehensive understanding of their nutritional status.

One limitation of this study is the lack of comparison with hybrid models such as LSTM-RF, LSTM-ARIMA, and LSTM-CNN. These models are highlighted in the literature for their strong predictive capabilities. Additionally, advanced spatio-temporal models like T-GCN (Temporal Graph Convolutional Network) are known for their ability to capture both spatial and temporal dependencies. Future research should explore these models to assess their performance against LSTM-FC and other baseline models. This could provide deeper insights into their applicability for time-series prediction and spatio-temporal analysis tasks. Furthermore, future research with the LSTM-FC model aims to first ensure access to a system equipped with a high-performance GPU to facilitate efficient hyperparameter tuning, enabling improved optimization and performance evaluation.