Data source.

This cross-sectional study utilized data from the Nepal Multiple Indicator Cluster Survey (NMICS) 2019. It was conducted by the Central Bureau of Statistics (CBS) of Nepal with technical and financial assistance from UNICEF [13]. NMICS applied a stratified cluster sampling design using 512 clusters across seven provinces in Nepal to implement an extensive survey of 12,800 households. It ensures careful and consistent information on maternal and child health, nutrition, and related demographic indicators. In the data set, 6,749 children aged below five years were eligible for anthropometric measurement. After a data cleaning exercise in which missing and miscoded values were removed, 6,416 valid cases were preserved for data analysis.

Variables and measurements.

Response variable creation: We used three WHO-standardized anthropometric z-scores: weight-for-age (WAZ), height-for-age (HAZ), and weight-for-height (WHZ). Following WHO guidelines, children with z-scores below −2 standard deviations were classified as underweight, stunted, and wasted, respectively [14]. A composite binary target variable ‘malnutrition’ was created, where a child was classified as malnourished if they exhibited any form of undernutrition (underweight or stunted or wasted). The rationale for using a composite outcome was primarily driven by class imbalance considerations. In the data set, 1,520 children (23.69%) were underweight, 2,077 children (32.37%) were stunted, and 763 children (11.89%) were wasted, individual conditions exhibited severe imbalance, with wasting alone being particularly problematic for classification modeling. Combining these into a composite outcome, the final data set (n = 6,416) has moderate class imbalance with 3,671 (57.2%) nourished and 2,745 (42.8%) malnourished children.

Stunting was the most prevalent condition, while wasting was least common – a pattern consistent with national epidemiological surveys in Nepal. The overlap structure (Table 1) demonstrates substantial co-occurrence of conditions with 1,397 children (51% of malnourished) presenting multiple concurrent indicators.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Overlap structure of malnutrition conditions in the dataset (n = 6,416).

https://doi.org/10.1371/journal.pone.0353969.t001

Domain-informed variable selection: Based on established literature on child malnutrition determinants and expert consultation, we selected 23 available features across six thematic domains aligned with UNICEF’s conceptual framework for malnutrition [10,11,15–17]:

1. Child care and supervision (n = 6): days left alone for more than one hour, left with a child (under 10) for more than one hour, took away privileges, explained why the behavior was wrong, shook the child
2. Child health and nutrition (n = 7): ever breastfed, currently breastfed, meal frequency, diarrhea in the last two weeks, fever in the last two weeks, cough in the last two weeks, vaccination record
3. Child characteristics (n = 4): child age, child sex, child disability, child able to pick up a small object
4. Maternal and household factors (n = 3): maternal education level, household wealth index, health insurance coverage
5. Geographic and environmental factors (n = 3): residence type (urban/rural), provinces, safe stool disposal practices

To prevent circularity between outcome definition and predictors, we adopted two methodological safeguards: (1) anthropometric z-scores (HAZ, WHZ, WAZ) and their components (height, weight) were excluded from the predictor set, ensuring no direct mathematical components of the outcome appeared among input features; and (2) child age was operationalized as a categorical variable with five broad categories (<1, 1, 2, 3, 4 years) rather than continuous age in months. This categorical encoding captures developmentally meaningful stages of early childhood while deliberately avoiding fine-grained age information that could introduce circularity with HAZ and WAZ calculations, which require precise monthly age for WHO reference comparisons. The detailed workflow is presented in Fig 1.

Pre-processing.

We adopted a domain-informed approach to handle missing data, recognizing that the absence of responses often carries meaningful information in survey data. Age-dependent questions (e.g., breastfeeding status for older children, discipline practices) where “not asked” occurred due to survey skip logic were encoded as −1, preserving this systematic missingness as an informative category. For variables with minimal missing data (<1%), such as recent illness indicators and health insurance status, mode imputation was applied. Response categories including “don’t know,” “no response,” and “missing/dk” were uniformly recoded as “not asked” to maintain consistency across the dataset. This approach retained the maximum sample size while preventing information loss from survey design features.

Features were encoded according to their measurement properties and theoretical relationships with malnutrition. Ordinal variables (wealth index: 1–5; mother’s education: 0–3; meal frequency: 0–7) preserved their natural ordering. Age-dependent behavioral variables were ternary-encoded (1/0/-1 for yes/no/not asked) to distinguish genuine responses from non-applicability. Binary variables (sex, residence type, insurance) received standard 0/1 encoding. Multi-category nominal variables were transformed as follows: stool disposal practices were collapsed into binary safe/unsafe categories based on WHO guidelines; vaccination status indicated possession of any documentation; and province was one-hot encoded into six indicators, with Bagmati as the reference category. This encoding strategy balanced model interpretability with capturing non-linear relationships while avoiding multicollinearity.

To address the moderate class imbalance, we applied Synthetic Minority Over-sampling Technique (SMOTE) [18] within each cross-validation fold. SMOTE was applied exclusively to training data after the train-test split to prevent data leakage, generating synthetic minority class examples through interpolation. Test data retained the original distribution to ensure evaluation metrics reflected real-world model performance. The preprocessing pipeline was applied uniformly across all 16 machine learning and deep learning models to ensure fair comparison. Feature scaling using StandardScaler (zero mean, unit variance) was additionally applied to models sensitive to feature magnitudes: linear methods (Logistic Regression, LDA), distance-based methods (KNN, SVM), and all neural networks (DNN, ResNet, Wide & Deep). Tree-based models (Decision Tree, Random Forest, Extra Trees), gradient boosting variants (XGBoost, LightGBM, CatBoost, Histogram Gradient Boosting, AdaBoost), and TabNet (which incorporates internal feature processing) used unscaled features, consistent with their algorithmic requirements. StandardScaler was fitted exclusively on training data and applied to test data to prevent data leakage. This standardized pipeline ensures that observed performance differences reflect genuine algorithmic capabilities rather than preprocessing inconsistencies.

Methodology.

Filter methods: Mutual information, Chi-square test, ANOVA F-statistic, Pearson correlation, and variance threshold to assess individual feature-target relationships.

Wrapper methods: Recursive feature elimination (RFE) with logistic regression and gradient boosting, plus Sequential forward selection maximizing cross-validated F1-score.

Embedded methods: Feature importance from random forest, gradient boosting, XGBoost, and L1-regularized logistic regression (LASSO).

Boruta algorithm: All-relevant feature selection using random forest wrapper (100 iterations, ) to identify features with statistically significant predictive power beyond random chance.

Ensemble aggregation: Features were ranked by average position across all 10 methods, with confirmed Boruta features receiving priority weighting.

Final features.

We selected 16 features based on Boruta confirmation supported by high ensemble rankings (average rank ) while capturing UNICEF’s multi-dimensional malnutrition framework [17]. This approach balanced statistical significance, domain relevance, and empirical performance (cross-validated ).

Selected features by category: Child care and supervision (away privileges, left alone), child health and nutrition (vaccination record, meal frequency, recent diarrhea, recent cough), child characteristics (child age), maternal and household factors (mother’s education, wealth index, health insurance), geographic and environmental factors (residence type, koshi, gandaki, karnali, sudurpaschim, safe stool disposal)

Features like recent diarrhea and sudurpaschim, though Boruta-rejected, were retained due to strong ensemble support and domain importance (WHO-recognized diarrhea-malnutrition pathway; geographic representation balance), as illustrated in Figs 2 and 3. The suitability of this multi-criteria feature selection strategy was empirically validated through a formal ablation analysis comparing our approach against strict Boruta-confirmed-only selection, with detailed results presented in S1 Appendix. The ablation analysis confirmed that retaining these domain-relevant features improved model performance across all evaluation metrics compared to strict Boruta-only selection.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Feature importance with their normalize score.

https://doi.org/10.1371/journal.pone.0353969.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Feature importance using Boruta method.

https://doi.org/10.1371/journal.pone.0353969.g003

Traditional machine learning models.

Logistic regression: Logistic regression is a widely-used statistical model for binary classification. It models the log-odds of the probability as a linear function of the predictors

The parameter vector can be estimated via maximum likelihood estimation, which then leads to the prediction of probability that an unseen case is malnourished.

K-nearest neighbor: K-nearest neighbor (KNN) is a non-parametric supervised machine learning algorithm that can be used for both classification and regression [19]. In the classification settings, KNN assigns the class that is most frequent among the k nearest neighbors of the observation,

where is the set of all possible classes, and is the k nearest neighbors of the observation with predictor x. The assignment is done by majority vote.

Decision tree: A decision tree classifier (DT) is a non-parametric supervised learning algorithm that recursively divides the input space into disjoint regions , each corresponding to a terminal node or leaf [20]. For any query point , the model assigns a predicted class label based on the majority class in the region that contains x. This is formally represented as

and the term denotes the empirical proportion of class c in region , computed as

Random forest: Random forest (RF) is an ensemble-based classification method that builds a collection of decision trees, each trained on a different bootstrap sample of the data and a random subset of features. This approach leverages the principle of aggregating the predictions of multiple uncorrelated models. During inference, each tree casts a “vote” for the predicted class, and the final output is determined by majority voting across the ensemble. The randomness introduced through both data sampling and feature selection reduces the correlation between individual trees, thereby enhancing model stability and minimizing overfitting. As the number of trees increases, the model generally becomes more robust and accurate, especially in high-dimensional or noisy datasets [21].

Extremely randomized trees: An ensemble learning method that constructs multiple decision trees using random subsets of features and random split points, rather than optimal splits. This additional randomness reduces variance more effectively than Random Forests while maintaining low bias, often improving generalization and reducing overfitting [22].

Linear discriminant analysis: Linear discriminant analysis (LDA) is a supervised classification technique that projects data onto a lower-dimensional space to maximize class separation and minimize variance within classes. LDA is similar to analysis of variance and regression analysis, as it expresses one response variable as a linear combination of other features to find the best separator among classes [23].

Support vector machine: Support vector machine (SVM) is a widely used supervised machine learning method for classification and regression analysis. It constructs a hyperplane that maximizes the distance to the closest data points from any class. A larger margin generally leads to better generalization performance, reducing the likelihood of overfitting and improving the model’s ability to perform well on unseen data. For linearly separable data, SVM solves the optimization problem

where are input vectors and are class labels.

In this study, we applied both a linear kernel (hyperparameter-tuned model), which assumes data can be separated in the original feature space, and the Radial Basis Function kernel (default model), a non-linear kernel that maps inputs into a higher-dimensional space using where is a kernel parameter controlling the spread [24].

Gradient boosting models.

Extreme gradient boosting: Extreme gradient boosting (XGBoost) is an advanced ensemble learning method that has gained significant popularity due to its high predictive performance across a wide range of applications, including classification and regression tasks. The method is an extension of the gradient boosting framework, which constructs models in a sequential manner. At each stage, a new decision tree is trained with the specific purpose of addressing the errors or residuals left by the preceding trees. This iterative refinement enables the model to progressively improve accuracy and reduce bias. In the case of binary classification, XGBoost typically relies on a logistic loss function, which provides an effective way of quantifying classification errors [25].

Light gradient boosting machine: A gradient boosting framework that employs histogram-based algorithms and Gradient-based One-Side Sampling (GOSS) with Exclusive Feature Bundling (EFB) to achieve faster training speed and higher efficiency. LightGBM uses leaf-wise tree growth strategy rather than level-wise, resulting in better accuracy with fewer iterations [26].

Categorical boosting: A gradient boosting algorithm specifically designed to handle categorical features efficiently using ordered target statistics and oblivious decision trees. CatBoost implements ordered boosting to avoid prediction shift and uses symmetric trees to reduce overfitting while maintaining high prediction speed [27].

Adaptive boosting: An ensemble meta-algorithm that combines multiple weak learners (typically shallow decision trees) by iteratively adjusting sample weights based on classification errors. Misclassified samples receive higher weights in subsequent iterations, forcing the algorithm to focus on difficult cases, resulting in a strong classifier from weak ones [28].

Histogram-based gradient boosting: A scikit-learn implementation of gradient boosting that uses histogram-based decision tree learning, binning continuous features into discrete bins. This approach significantly reduces memory consumption and computational cost while enabling efficient handling of missing values and support for monotonic constraints.

Deep learning models.

Deep neural network: A feedforward artificial neural network with multiple hidden layers between input and output layers, utilizing batch normalization, dropout regularization, and LeakyReLU activation functions to learn hierarchical feature representations. The deep architecture enables automatic feature extraction and non-linear pattern recognition in complex datasets [29].

ResNet: A deep neural network architecture that employs residual connections (skip connections) to address the vanishing gradient problem in very deep networks. These shortcut connections allow gradients to flow directly through the network, enabling training of substantially deeper models while maintaining or improving performance. ResNet-style models often perform well compared to other neural approaches and serve as a useful baseline in tabular deep learning research [30].

Wide and deep network: A hybrid neural network architecture that combines a wide linear model (for memorization of feature interactions) with a deep neural network (for generalization through feature learning). This dual-path structure enables both memorization of specific patterns and generalization to unseen data, particularly effective for recommendation systems and tabular data [31].

TabNet: TabNet is a deep learning architecture specifically tailored for tabular (structured) data, developed by Google Cloud AI. It introduces a novel approach by using sparse feature selection through sequential attention mechanisms. It allows the model to focus on the most informative features at each decision step. This adaptive feature selection process improves both learning efficiency and model interpretability. At each step, TabNet learns a feature mask that determines which subset of input variables should be emphasized, enabling it to reason selectively across multiple layers. The architecture supports local interpretability by showing how features influence individual predictions, and global interpretability by assessing the overall contribution of each feature across the dataset. This combination of performance and explainability makes TabNet particularly effective for structured data modeling tasks [32].