Evaluating LIME and SHAP in Explaining Malnutrition Classiﬁcation in

Children Under Five

Nuru Nabuuso

Department of Engineering, Universitat Pompeu Fabra, Roc Boronat 138, Barcelona 08018, Spain

Keywords:

Machine Learning Methods, Classiﬁcation, XGBoost, Feature Selection, Explainable AI.

Abstract:

Malnutrition in children under ﬁve is a signiﬁcant public health issue in Uganda, with severe impacts on devel-

opment and mortality. This paper explores machine learning (ML) models—Support Vector Machines (SVM),

eXtreme Gradient Boosting (XGBoost), and Artiﬁcial Neural Networks (ANNs) — to predict malnutrition,

and reports that XGBoost shows highest predictive accuracy. While the ﬁndings on XGBoost employed global

model interpretation through feature importance based on permutations, we also introduce SHapley Additive

exPlanations (SHAP) for both local and global interpretations. We follow with a focus on SHAP summary

plots and bar charts to evaluate feature importance globally. In addition, we report on the comparison be-

tween SHAP and Local Interpretable Model-agnostic Explanations (LIME) to analyze the consistency of local

explanations provided by both techniques. By contrasting LIME and SHAP, we advance the alignment be-

tween local and global interpretations in the context of XGBoost predictions. This comparison highlights the

strengths and limitations of each method. Our ﬁndings aim to enhance the transparency of ML models and

improve decision-making in child health interventions, providing signiﬁcant insights into public health and

ML interpretability.

1 INTRODUCTION

Malnutrition among children under ﬁve years of age

remains a signiﬁcant public health issue, particularly

in low-resource settings such as Uganda (Kikafunda

et al., 1998). Early detection and intervention are

critical for improving outcomes, but identifying at-

risk children is often challenging due to the com-

plex interplay of factors that contribute to malnu-

trition (Sermet-Gaudelus et al., 2000). ML tech-

niques have shown great promise in predictive mod-

eling for health outcomes, offering the potential to

enhance early diagnosis and provide tailored inter-

ventions (Talukder and Ahammed, 2020; Islam et al.,

2022; Bitew et al., 2022; Gadekallu et al., 2021). We

report here on the application of ML models, includ-

ing SVM, XGBoost, and ANNs to predict malnutrition

in children under ﬁve.

While Artiﬁcial Intelligence and ML are penetrat-

ing many ﬁelds, there is an urgent and pressing need

not only to achieve high accuracy but also to achieve

explainability. Despite the potential of ML models,

relatively little research has focused on explaining the

predictions of models speciﬁcally designed to iden-

tify malnutrition in children under ﬁve (Talukder and

Ahammed, 2020). In health contexts, where decisions

can have profound effects on patient outcomes, in-

terpretability is just as important as predictive accu-

racy (ElShawi et al., 2021). Explainability is crucial

for transparency (now required in some regions by

legislation) and for trust (essential for user adoption).

Clinicians, public health ofﬁcials, and policy makers

require models that not only provide accurate predic-

tions but also offer clear and understandable insights

into the factors driving those predictions. To address

this need, we employ feature importance based on

permutations to generate global explanations for the

model, offering an overall view of how features in-

ﬂuence predictions. However, this method may not

capture the complete dynamics of feature interactions

in complex models. LIME and SHAP are arguably

the most used approaches to gain insights into the

outputs produced by classiﬁers and generate expla-

nations in human-understandable terms. This study

aims to provide local and global explanations using

SHAP (Lundberg and Lee, 2017).

SHAP offers a more theoretically grounded ap-

proach by assigning importance scores to features

based on their contribution to the model’s predictions.

To assess global feature importance, we use SHAP

Nabuuso, N.

Evaluating LIME and SHAP in Explaining Malnutrition Classiﬁcation in Children Under Five.

DOI: 10.5220/0013186500003905

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 291-298

ISBN: 978-989-758-730-6; ISSN: 2184-4313

291

summary plots and bar charts. We compare SHAP’s

results with permutation-based feature importance to

evaluate whether global interpretations remain con-

sistent across these methods. Sometimes LIME and

SHAP concur in their conclusions about a classiﬁer’s

decision that has been deemed the most suitable for

some applications. Sometimes this is not the case.

Therefore, we contrast here SHAP with LIME for lo-

cal interpretations (Ribeiro et al., 2016). LIME gener-

ates locally interpretable explanations by approximat-

ing the model’s behavior in the vicinity of individ-

ual predictions, while SHAP provides consistent lo-

cal explanations by leveraging cooperative game the-

ory. By comparing both local and global explana-

tions, we aim to assess the consistency and effective-

ness of these interpretability techniques in enhancing

model transparency and improving decision-making

in child health interventions.

We will ﬁrst report on the application of ML to

predict malnutrition. Among the models we tested,

the XGBoost classiﬁer achieved the highest accuracy

in predicting whether a child is stunted, wasted, or

underweight. We, therefore, further explore the ex-

plainability of XGBoost results through interpretabil-

ity techniques to enhance understanding and trust in

the model’s predictions (Ribeiro et al., 2016). There-

fore, this paper aims to also evaluate the consistency

of XGBoost results through a thorough assessment of

its interpretability using SHAP and LIME.

2 RELATED RESEARCH

We proceed to review some studies in the context of

health where, at least LIME and SHAP have been

used for explainability (among perhaps some other

approaches). (Kumar et al., 2024) explored the ap-

plication of various ML models for detecting ane-

mia and predicting its severity. The study analyzed

a dataset of 364 individuals, using Logistic Regres-

sion, K-Nearest Neighbors (KNN), SVM, Decision

Tree, and Random Forest. (Pedregosa et al., 2011)’s

study applies two implementations of hyperparam-

eter ﬁnding: GridSearchCV and RandomSearchCV.

Additionally, the study evaluated boosting techniques

like AdaBoost, Gradient Boosting, CatBoost, and

XGBoost. To understand the model’s predictions bet-

ter, they employed LIME and SHAP. Among the

models used, Random Forest gave the highest accu-

racy of 89.04%. The accuracy was 86.30% for the

Decision Tree classiﬁer, 87.67% for logistic regres-

sion, 78.08% for KNN and 87.67% for SVM. Making

Random Forest the best choice for such a data set.

(Aldughayﬁq et al., 2023) explored LIME and

SHAP to generate local and global explanations

for a deep learning model based on inceptive V3

architecture trained on retinoblastoma and non-

retinoblastoma fungus images. Since deep learning

models are considered black-box models, they ap-

plied LIME and SHAP to generate explanations on

the validation and test sets. Their results showed

that LIME and SHAP provided valuable insights and

showed areas or parts of the images that contributed

to the models’ predictions both locally and globally.

In that research, SHAP provided more accurate results

and provided effective explanations in identifying the

important sections of the images.

LIME and SHAP explanations have also been ex-

tensively used in analyzing Electronic Health Records

(EHRs) (Di Martino et al., 2023), where LIME, SHAP

and Scoped Rules (Ribeiro et al., 2018) are applied

to compute feature importance for ML predictions.

These explainability techniques generated top fea-

tures, offering deeper insights into the model’s results.

In that study, three XAI methods were employed to

demonstrate the effectiveness of explainable conclu-

sions in ML models and provide data interpretability

for large-scale EHR data. Speciﬁcally, ML models

were applied alongside XAI methods to study lung

cancer mortality.

3 SUITABLE CLASSIFIER

In this section, we report in our evaluation of three

ML approaches to predict malnutrition in children

under ﬁve. We use data from the 2016 Uganda De-

mographic and Health Survey (UDHS)https://www.

dhsprogram.com/data/dataset admin/login main.cfm

for our study. The Uganda Bureau of Statistics

(UBOS) implemented the 2016 UDHS and covers

household and respondent characteristics. The

dataset consists of 5379 records and includes history

of all women and children health born in the last

5 years prior to the survey (our unit of analysis)

with parental or guardian consent. The three ML

classiﬁcation models we used are as follows.

XGBoost: It is an efﬁcient implementation of the gra-

dient boosting algorithm designed for both classi-

ﬁcation and regression tasks. It builds an ensem-

ble of decision trees sequentially, where each tree

attempts to correct the errors of the previous ones.

SVM: It works by ﬁnding the hyperplane that best

separates data into different classes after mapping

to a higher dimensional space.

ANN s: It is a computational model, consisting of in-

terconnected nodes (neurons) organized in layers.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

292

These networks are particularly adept at capturing

complex patterns in large datasets through the use

of non-linear activation functions.

3.1 Study Variables and Measurements

The classes of interest are stunted (that involves the

ratio of height to age — h/a), wasted (for the ratio

of weight to height — w/h) and underweight (consid-

ering the ratio of weight to age — w/a). We used

Z-scores of anthropometric measurements to evalu-

ate the nutritional status for the children. We used

the World Health Organisation (WHO) AnthroPlus

software to compute the Z-scores (WHO, 2007). The

WHO deﬁnes the Z-scores for the class labels as dis-

placement (in proportion to the corresponding stan-

dard deviation σ) from the corresponding mean value

µ:

stunted: height-to-age < −2 × σ

h/a

+ µ

h/a

;

wasted: weight-to-height < −2 × σ

w/h

+ µ

w/h

and

underweight: weight-to-age < −2 × σ

w/a

+ µ

w/a

Severely stunted, wasted and underweight are those

children whose height-to-age, weight-to-height or

weight-to-age Z-score are below minus 3 (-3) stan-

dard deviations from the corresponding median. The

classes were binary coded as 1 for stunted, wasted and

underweight if the standard was met, else they were

coded as 0.

3.2 Data Preprocessing and Feature

Selection

We encoded the data into numerical values using one-

hot encoding. We removed data noise and inconsis-

tencies and used box plots to remove outliers too.

We replaced missing values with the mean or mode

depending on the data structure and used linear re-

gression to predict missing values for anthropome-

try measurements that were used to compute the Z-

Scores. We computed the correlation matrix to ﬁnd

the strength of association between independent vari-

ables using an absolute value of 0.6 as the thresh-

old for retaining a variable. If two variables were

found to be correlated, we dropped the variable with a

lower correlation coefﬁcient value to the target class.

We also computed multiple correlation coefﬁcients to

check whether more than two variables are correlated.

We employed the feature permutation method to eval-

uate the importance of features for the models. This

method consists of randomly shufﬂing the values of

a speciﬁc feature to measure the impact on classiﬁca-

tion though it is computationally expensive.

The imbalanced dataset challenge was addressed

by using SMOTE technique which randomly in-

creases the minority class examples thereby pre-

venting over-ﬁtting. For this study we used strat-

iﬁed cross-validation. In K-Fold cross-validation,

the dataset is split into K smaller sets or “folds”.

The model is trained on K-1 folds and tested on

the remaining fold. The stratiﬁed approach en-

sured that the evaluation metrics were reliable, even

with skewed class distributions, and helped in ﬁne-

tuning the model parameters by providing insights

into its performance on various data partitions. The

cross-validation process not only improved the overall

model accuracy but also helped mitigate over-ﬁtting,

making the models more generalizable to unseen

data.By combining stratiﬁed K-Fold cross-validation

with the SMOTE technique we ensured that the mod-

els were both accurate and resilient to the dataset.

3.3 Model Evaluation and Performance

Comparison

We applied several evaluation metrics to assess model

performance. A confusion matrix was used to de-

termine True Positives (TP), True Negatives (TN),

False Positives (FP), and False Negatives (FN). This

allowed for the calculation of key metrics such as Ac-

curacy, Sensitivity, and Speciﬁcity. All models per-

formed well with variations depending on the tuning

technique applied to them. The SVM gave its best per-

formance when the RBF kernel was used, see Table 1.

This kernel out-performed other kernels like Poly,

Linear and Sigmoid. We applied GridsearchCV,

and this hyper-parametrization optimized all models,

resulting in XGBoost displaying the best respective

performance (see Table 2).

Table 1: SVM kernel accuracy (as percentage).

Kernel type stunted underweight wasted

RBF 64.2 89.0 94.3

Linear 51.0 58.6 59.2

Poly 49.6 50.2 50.8

Sigmoid 50.3 51.2 62.8

Table 2 shows superior results by XGBoost classi-

ﬁer outperforming SVM and ANN across all classes.

Table 2: Accuracy (as percentage with 95% conﬁdence

interval).

Classiﬁer stunted underweight wasted

SVM 64.21 ± 1.02 89.02 ± 0.43 94.35 ± 0.43

ANN 62.52 ± 0.43 62.95 ± 0.58 61.98 ± 0.82

XGBoost 74.35 ± 0.82 95.67 ± 0.56 98.17 ± 0.43

Evaluating LIME and SHAP in Explaining Malnutrition Classiﬁcation in Children Under Five

293

4 EXPLAINABLE AI METHODS

4.1 Hypotheses

We hypothesise that LIME and SHAP provide more

consistent and accurate local and global interpreta-

tions of XGBoost model predictions for malnutrition

in children under ﬁve years.

4.2 XAI Methods

LIME provides local interpretation for black-box

models, such as XGBoost, by approximating the com-

plex model f locally around a speciﬁc instance x us-

ing a simpler interpretable model g (Zhang et al.,

2019; Lee et al., 2019). The objective is to minimize a

loss function L that measures how well g matches f ’s

predictions for instances sampled around x, deﬁned

as:

L(g) =

∑

∈Z

) · ( f (x

) − g(x

))

where Z is the set of sampled instances, π

) is the

proximity weight of x

relative to x, f (x

) is the pre-

diction from f , and g(x

) is the prediction from g. A

regularization term Ω(g), such as the number of fea-

tures used, penalizes complexity. The ﬁnal objective

is to minimize the total loss, combining prediction er-

ror and complexity.

SHAP explains feature contributions to predic-

tions using Shapley values from cooperative game

theory (Rodr

ıguez-P

erez and Bajorath, 2019). The

relevance of a feature i is a player contribution φ

the ﬁnal prediction deﬁned as:

∑

S⊆Z\{i}

|S|! (|Z| − |S| − 1)!

|Z|!

[ f (S ∪ {i}) − f (S)],

where Z is the set of all features, S is a subset of fea-

tures excluding i, f (S) is the prediction based on S,

and f (S ∪ {i}) is the prediction when i is added to S.

The weight ensures a fair contribution based on sub-

sets’ sizes.

Feature Permutations assess feature importance

by measuring the performance drop when feature val-

ues are randomly permuted (Casalicchio et al., 2019).

Permuting values of inﬂuential features signiﬁcantly

erodes model performance, while less important fea-

tures have minimal impact. However, this method

can underestimate the importance of highly correlated

features. We use the above XAI methods to explain

the XGBoost model globally and locally.

5 EXPERIMENTS

We independently ﬁtted each class using the XGBoost

classiﬁer because of its superior performance in pre-

dicting malnutrition. To explain the XGBoost model

globally, we used the feature permutation method and

SHAP method. Both approaches provide insights into

the overall behavior of the model, identifying the most

inﬂuential features across the dataset.

For feature permutations, we created a copy of the

dataset where the values of each feature were ran-

domly permuted, breaking its relationship with the

target variable while keeping other features intact.

The XGBoost model then made predictions on the

permuted datasets, and we compared the performance

metrics with those from the original dataset. The drop

in performance for each feature indicated its impor-

tance; more signiﬁcant drops signiﬁed greater impor-

tance. We performed multiple iterations of the per-

mutation process to obtain stable estimates of feature

importance, which we averaged to summarize each

feature’s impact.

Conversely, we computed SHAP values using the

SHAP library installed in our Python environment.

These values provided an alternative perspective on

feature importance, complementing the insights de-

rived from feature permutations.

6 INTERPRETATION AND

DISCUSSION

The comparison of feature importance results from

feature permutations and SHAP revealed consistent

ﬁndings in globally explaining the XGBoost model.

Both methods identiﬁed the same top three fea-

tures—size of child at birth, partner’s age, and age of

household head—for predicting stunted and wasted.

For the class underweight, the top two features were

consistent, while the rankings of other features varied

across the three malnutrition indicators, as shown in

Table 3.

Table 3: Table showing the top three features for the differ-

ent methods.

Class Top Permutation Features Top SHAP Features

stunted Size of child at birth Size of child at birth

Partner’s age Partner’s age

Age of household Head Age of household Head

underweight Size of child at birth Size of child at birth

Age of household Head Age of household Head

Number of children Partner’s age

wasted Size of child at birth Size of child at birth

Place of delivery Number of U5 in household

Number of U5 in household Place of delivery

This consistency in the top features highlights

their strong and stable inﬂuence on the model’s pre-

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

294

dictions. However, the variability in other fea-

ture rankings suggests that these may be context-

dependent or inﬂuenced by interactions with the con-

sistently ranked features. Such insights are crucial for

understanding the dynamics of malnutrition and de-

signing targeted interventions. The SHAP beeswarm

plots (Figure 1 and Figure 2) provide additional con-

text by illustrating how individual feature values cor-

relate with model predictions. For instance, larger

sizes at birth are associated with lower risks of stunt-

ing, while older partners reﬂect better household re-

sources. These visualizations complement the results

of permutation importance, as they show both the

magnitude and direction of feature impacts across dif-

ferent contexts. By comparing SHAP’s results with

permutation importance, the analysis validates the

strength and consistency of the identiﬁed features.

Both methods conﬁrm the importance of the same

features and provide a comprehensive view of their ef-

fects on model predictions. Together, they strengthen

the interpretation of the XGBoost classiﬁer by offer-

ing robust global insights into feature importance, as

illustrated by the referenced tables and ﬁgures (refer

to Figure 1).

7 LOCAL EXPLANATION USING

LIME AND SHAP

To apply LIME, we randomly selected instance 80

within the class stunted, the class wasted, and the

class underweight and generated local explanations

using both LIME and SHAP. We compared their fea-

ture importance, magnitudes of importance, and qual-

itative differences in their explanations.

For feature importance, we extracted the top N

features identiﬁed by LIME and SHAP as contributing

most to the model’s prediction, comparing their rank-

ings to identify similarities or differences. We com-

pared the magnitude of feature contributions using the

absolute values of the contributions provided by both

methods. We assessed whether they assign similar or

different levels of importance to key features.

Finally, we explored the qualitative differences be-

tween LIME and SHAP explanations, focusing on the

localized nature of LIME’s explanations compared to

the local consistency of SHAP. This reveals insights

unique to each method. Since the visual represen-

tations for LIME and SHAP facilitate comparison,

we visually inspected the agreement or divergence in

their explanations.

7.1 Interpretation

7.1.1 Class Stunting

SHAP and LIME provided different feature impor-

tance to explain the same local instance. LIME iden-

tiﬁed (1) source of drinking water, (2) partners age

and (3) age of household head as the top features that

impact stunted. However, SHAP identiﬁed, (1) size of

child at birth, (2) partners age, (3) number of U5 in

household as the top three features (see Figure 4). Al-

though the top three features have different rankings,

partners age was ranked second for the two methods,

positively impacting the prediction in both cases. This

placing means that at some point, these two methods

agreed.

7.1.2 Class Underweight

LIME identiﬁed the top four features as (1) number of

antenatal visits, (2) partner’s age, (3) source of drink-

ing water and (4) size of child at birth in their order of

importance. Meanwhile, SHAP identiﬁed (1) number

of antenatal visits, (2) size of child at birth, (3) Num

of children and (4) partners age also in their impor-

tance order. These two methods shared the topmost

features, and they all showed that it negatively im-

pacted underweight (see Figure 5). In common, the

two methods shared three features in the ranked top 4

and all shared features agreed on how they impacted

the output; that is, all the same feature negatively im-

pacted the prediction.

7.1.3 Class Wasting

When explaining the local instance for wasted, LIME

identiﬁed (1) number of U5 in household, (2) size of

child at birth, (3) place of delivery, and (4) time to

water source as the top-ranked features. SHAP identi-

ﬁed (1) number of U5 in household, (2) size of child at

birth, (3) source of drinking water and (4) time to wa-

ter source. LIME and SHAP identiﬁed the same top

two features and the 4th feature coincides across the

two methods. These features that were similar to both

methods also had the same ranking across as well as

the same impact on the prediction. That is, all three

features agreed on how they are positively or nega-

tively impacting the prediction (see Figure 6).

The agreement indicates that these features are

likely to contribute to the model’s decision-making

process signiﬁcantly. The fact that the three features

are not only top-ranked but also exhibit similar inﬂu-

ence, whether positive or negative, points to a coher-

ent narrative about the model’s behavior for this spe-

ciﬁc prediction. It also suggests that the model cap-

Evaluating LIME and SHAP in Explaining Malnutrition Classiﬁcation in Children Under Five

295

(a) Permutation feature importance plot showing the impact

of each feature on predicting stunted.

(b) SHAP beeswarm plot showing the impact of each fea-

ture on predicting stunted.

Figure 1: Comparisons of permutation feature importance and SHAP results for the class stunted.

(a) Permutation feature importance plot showing the impact

of each feature on predicting underweight.

(b) SHAP beeswarm plot showing the impact of each fea-

ture on predicting underweight.

Figure 2: Comparisons of permutation feature importance and SHAP results for the class underweight.

(a) Permutation feature importance plot showing the impact

of each feature on predicting wasted.

(b) SHAP beeswarm plot showing the impact of each fea-

ture on predicting wasted.

Figure 3: Comparisons of feature permutation’s importance and SHAP results for the class wasted.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

296

(a) LIME instance explanation for the class stunted. (b) SHAP instance explanation for the

classstunted.

Figure 4: Comparisons of LIME and SHAP results for the class stunted.

(a) LIME instance explanation for the class underweight. (b) SHAP instance explanation for the class under-

weight.

Figure 5: Comparisons of LIME and SHAP results for the class underweight.

(a) LIME instance explanation for the class wasted. (b) SHAP instance explanation for the class wasted.

Figure 6: Comparisons of LIME and SHAP results for the class wasted.

tures meaningful patterns, not just artifacts of one par-

ticular interpretability method. This consistency can

enhance stakeholder conﬁdence in the model’s pre-

dictions and provide actionable and trusted insights

for decision-making.

8 CONCLUSIONS

We compared SHAP and LIME to evaluate the con-

sistency of local explanations provided by both meth-

ods and to compare globally identiﬁed feature im-

portance using SHAP values and feature permutation.

The convergence of insights from both feature permu-

tation and SHAP in explaining the predictions of the

XGBoost model for the classes stunted, wasted, and

underweight highlights a robust understanding of the

underlying factors inﬂuencing these classes.

The alignment between these methods enhances

conﬁdence in the identiﬁed features, demonstrating

their consistent contribution to the model’s decisions.

Similarly, the agreement between LIME and SHAP

Evaluating LIME and SHAP in Explaining Malnutrition Classiﬁcation in Children Under Five

297

underscores a robust interpretation of the model’s

workings, further validating the relevance of the iden-

tiﬁed features as reﬂections of underlying patterns

rather than artifacts of speciﬁc methods.

Additionally, the commons feature’s im-

pact—whether positive or negative—reinforces

the reliability of these features in inﬂuencing the

model’s outcomes. This convergence of results is

signiﬁcant for practitioners, indicating that both inter-

pretability methods provide a similar understanding

of the model, enabling more precise insights for

decision-making.

ACKNOWLEDGMENTS

The author acknowledges the contribution of Ms.

Irene Wanyana, Dr. Isunju JohnBosco and Dr. Kiberu

Vincent who supervised part of this research while the

author was completing a Master Program at Makerere

University and Vladimir Estivill-Castro as current ad-

visor during the author’s PhD program.

REFERENCES

Aldughayﬁq, B., Ashfaq, F., Jhanjhi, N., and Humayun, M.

(2023). Explainable ai for retinoblastoma diagnosis:

interpreting deep learning models with lime and shap.

Diagnostics, 13(11):1932.

Bitew, F. H., Sparks, C. S., and Nyarko, S. H. (2022). Ma-

chine learning algorithms for predicting undernutri-

tion among under-ﬁve children in ethiopia. Public

health nutrition, 25(2):269–280.

Casalicchio, G., Molnar, C., and Bischl, B. (2019). Visu-

alizing the feature importance for black box models.

In Proc. P. I Machine Learning and Knowledge Dis-

covery in Databases: European Conf., ECML PKDD

2018, pages 655–670. Springer.

Di Martino, F., Delmastro, F., and Dolciotti, C. (2023). Ex-

plainable ai for malnutrition risk prediction from m-

health and clinical data. Smart Health, 30:100429.

ElShawi, R., Sherif, Y., Al-Mallah, M., and Sakr, S. (2021).

Interpretability in healthcare: A comparative study

of local machine learning interpretability techniques.

Computational Intelligence, 37(4):1633–1650.

Gadekallu, T. R., Iwendi, C., Wei, C., and Xin, Q. (2021).

Identiﬁcation of malnutrition and prediction of bmi

from facial images using real-time image processing

and machine learning. IET Image Process, 16:647–

658.

Islam, M. M., Rahman, M. J., Islam, M. M., Roy, D. C.,

Ahmed, N. F., Hussain, S., Amanullah, M., Abedin,

M. M., and Maniruzzaman, M. (2022). Application

of machine learning based algorithm for prediction of

malnutrition among women in bangladesh. Int. J. of

Cognitive Computing in Engineering, 3:46–57.

Kikafunda, J. K., Walker, A. F., Collett, D., and Tumwine,

J. K. (1998). Risk factors for early childhood malnu-

trition in Uganda. Pediatrics, 102(4):e45–e45.

Kumar, A., Chirag, Y., Kodipalli, A., and Rao, T. (2024).

Anemia detection and severity prediction using clas-

siﬁcation algorithms with optimized hyperparameters,

boosting techniques and xai. In 2024 5th Int. Conf. for

Emerging Technology (INCET), pages 1–5. IEEE.

Lee, E., Braines, D., Stifﬂer, M., Hudler, A., and Harborne,

D. (2019). Developing the sensitivity of lime for bet-

ter machine learning explanation. In Artiﬁcial intelli-

gence and machine learning for multi-domain opera-

tions applications, v. 11006, pages 349–356. SPIE.

Lundberg, S. M. and Lee, S.-I. (2017). A uniﬁed ap-

proach to interpreting model predictions. In Proc. 31st

Int. Conf. on Neural Information Processing Systems,

NIPS, page 4768–4777, Red Hook, NY, USA. Curran

Associates.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer,

P., Weiss, R., Dubourg, V., et al. (2011). Scikit-

learn: Machine learning in python. Journal of ma-

chine learning research, 12(Oct):2825–2830.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2016). ”Why

should I trust you?”: Explaining the predictions of any

classiﬁer. In Proc. 22nd ACM SIGKDD Int. Conf. on

Knowledge Discovery and Data Mining, KDD, page

1135–1144, NY, USA. ACM.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2018). An-

chors: High-precision model-agnostic explanations.

In Proc. 32nd AAAI Conf. on artiﬁcial intelligence.

AAAI Press.

Rodr

ıguez-P

erez, R. and Bajorath, J. (2019). Interpreta-

tion of compound activity predictions from complex

machine learning models using local approximations

and shapley values. Journal of medicinal chemistry,

63(16):8761–8777.

Sermet-Gaudelus, I., Poisson-Salomon, A.-S., Colomb, V.,

Brusset, M.-C., Mosser, F., Berrier, F., and Ricour, C.

(2000). Simple pediatric nutritional risk score to iden-

tify children at risk of malnutrition. The American

journal of clinical nutrition, 72(1):64–70.

Talukder, A. and Ahammed, B. (2020). Machine learning

algorithms for predicting malnutrition among under-

ﬁve children in bangladesh. Nutrition, 78:110861.

WHO (2007). World health organisation application

tools. https://www.who.int/tools/growth-reference-

data-for-5to19-years/application-tools. Accessed:

2023-11-06.

Zhang, Y., Song, K., Sun, Y., Tan, S., and Udell, M. (2019).

” why should you trust my explanation?” understand-

ing uncertainty in lime explanations. arXiv preprint

arXiv:1904.12991.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

298