Can Zero-Shot LLMs Predict Child Malnutrition? A Fairness and Temporal Robustness Study

Md Ahshanul Haque; Muhammad Ashad Kabir

REVIEW 4 major objections 5 minor 14 references

Zero-shot language models can predict child stunting from survey records about as well as a supervised classifier, catching more true cases while over-flagging disadvantaged groups.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 13:57 UTC pith:TWP6RWR7

load-bearing objection Interesting exploratory result but zero-shot claim not yet established due to prompt tuning and contamination concerns; worth a serious referee if the authors fix the evaluation pipeline. the 4 major comments →

arxiv 2607.29082 v1 pith:TWP6RWR7 submitted 2026-07-31 cs.CL

Can Zero-Shot LLMs Predict Child Malnutrition? A Fairness and Temporal Robustness Study

Muhammad Ashad Kabir , Md Ahshanul Haque This is my paper

classification cs.CL

keywords child malnutritionzero-shot learninglarge language modelsstunting predictionfairnesstemporal robustnessGPT-4o-minidemographic health surveys

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper asks whether a pretrained large language model, used with no task-specific training, can predict child stunting from household-survey records. The authors turn each child's maternal, child, and household characteristics into a simple text prompt and ask GPT-4o-mini to output 0 or 1. They find that the zero-shot model reaches balanced accuracy comparable to a random forest baseline (58% vs 57%) but with a very different trade-off: it catches roughly three-quarters of stunted children (77.5% sensitivity) while also labelling many non-stunted children as stunted (38.6% specificity). The model is stable across survey waves from 2007 to 2022 and fair across child sexes, but it over-predicts stunting for rural and poor households, an important caution for real-world use.

Core claim

On the paper's own terms, the central finding is that zero-shot inference with GPT-4o-mini produces clinically meaningful stunting predictions from structured survey data without any fine-tuning: it matches the supervised baseline on balanced accuracy and substantially outperforms it on sensitivity, at the cost of specificity. The authors show that this behaviour is stable across sex groups and survey years, and they document a fairness gap in which children from rural areas and the poorest wealth categories are overwhelmingly flagged as stunted, with 100% sensitivity and 0% specificity for the poorest group. Excluding the wealth index from the prompt does not remove this pattern, suggesting

What carries the argument

The method's engine is prompt-based serialization: each child's feature–value pairs are written as a list-style text record, the model is instructed to act as a clinical classifier and answer with a single token, and probabilities are derived from the log-probabilities of the tokens 0 and 1. This lets a frozen pretrained LLM act as a zero-shot predictor. A random forest with balanced class weights serves as the supervised comparison, and fairness is measured by stratification on child sex, residence, and wealth index; temporal robustness is measured by comparing performance across five BDHS survey rounds.

Load-bearing premise

The whole evaluation assumes the model has not been contaminated: that GPT-4o-mini's pretraining has not memorized the specific survey waves or stunting patterns being tested, and that the prompt template, which the authors say was iteratively refined through empirical validation, was not tuned on the evaluation records.

What would settle it

Take the same 17,106 survey records and randomly shuffle the feature values within each record, keeping the stunting labels unchanged; then re-run the zero-shot prompts. If the LLM still produces the same sensitivity and specificity (or near-identical subgroup disparities), it is responding to prompt base rates rather than the actual features, and the fairness numbers would reflect the prompt design rather than the data. Alternatively, hold out a proper prompt-development split and show that balanced accuracy holds on records never seen during prompt iteration.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

If correct, zero-shot LLMs can serve as a no-training baseline for malnutrition screening from survey data, with a recall-oriented bias that may be useful for triage.
The same prompt-based pipeline could be applied to other demographic health surveys without retraining, as long as fairness gaps are audited.
The stable performance across 2007–2022 suggests the model's predictions are not driven by a single survey year's peculiarities.
The sensitivity/specificity asymmetry means deployment decisions require choosing whether false positives (over-referral) or false negatives (missed cases) are costlier.
The wealth and residence disparities imply that fair deployment would need calibration or threshold adjustment per subgroup.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The near-perfect sensitivity and zero specificity for the poorest group suggests the LLM may be using socioeconomic indicators as a stunting heuristic; since removing wealth index barely changes the pattern, the heuristic is likely distributed across correlated features such as education, sanitation, and delivery care.
A plausible editorial worry is contamination: the model may have encountered stunting statistics or DHS-style descriptions during pretraining, which the paper does not test; if so, the 'zero-shot' claim would be weakened.
The authors' choice of balanced accuracy as the headline metric masks the large calibration difference; a reader should look at the full sensitivity/specificity pair before drawing conclusions.
A concrete next test would be to apply the same prompt template to data from another country or a synthetic dataset with rearranged feature values; if the model's predictions depend mostly on base rates rather than the actual record, the current performance numbers would need to be reinterpreted.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Interesting exploratory result but zero-shot claim not yet established due to prompt tuning and contamination concerns; worth a serious referee if the authors fix the evaluation pipeline.

read the letter

First, the new result: this is the first test of a zero-shot LLM (GPT-4o-mini) on BDHS stunting prediction, with fairness and temporal breakdowns. That is a legitimate extension of TabLLM-style serialization to a public-health dataset, and the paper earns credit for reporting subgroup metrics and for re-running the poorest-wealth analysis without the wealth feature. The finding that balanced accuracy roughly matches a random forest (58 vs 57) while AUROC is lower (0.632 vs 0.685) is honestly presented, and the high-sensitivity/low-specificity trade-off is discussed.

The soft spots are real, though. The 'zero-shot' label is the main one. Section 2.2 says the prompt was 'iteratively refined through empirical validation' but gives no development/evaluation split. That is model selection on the test set. Alongside it, there is no contamination check for GPT-4o-mini, which is a real risk because DHS reports and derived text are online. So the headline claim—that the model matches a supervised baseline without task-specific training—is not established by this paper. The balanced-accuracy parity is threshold-dependent; the LLM's AUROC is below the RF's, so the equality is a single operating point, not evidence of comparable overall skill. Also, the temporal rows in Table 4 sum to 7,052, not the stated 17,106; that is a documentation gap that needs fixing. And the RF baseline has no error bars, so the 1-point balanced-accuracy gap is within noise.

These are addressable, not fatal. A proper zero-shot evaluation requires a fixed prompt (or a prompt-development split), a contamination probe (e.g., testing on a held-out dataset not plausibly in pretraining), and a consistent sample accounting. The paper is exploratory and reads like a workshop paper, which is fine at that venue.

I'd send it to peer review: the question matters, and the paper raises a caveat—fairness disparities across wealth and residence—that is worth putting on the record. But I'd ask the authors to clearly state the prompt tuning and contamination limitations, or even better, to run the analysis with a held-out prompt split. If they cannot rule out contamination, the abstract's 'zero-shot' phrase should be softened to 'prompt-based without gradient training.'

Useful for a reading group on LLM evaluation pitfalls, but not something I'd cite in my own work until the evaluation is cleaned up.

Referee Report

4 major / 5 minor

Summary. The paper evaluates GPT-4o-mini as a zero-shot predictor of child stunting from Bangladesh Demographic and Health Survey records (n = 17,106). Each child's features are serialized into a list-style prompt, and the LLM's token-level log-probabilities for '0' and '1' are used for classification. The model is compared with a random forest baseline and further analyzed by child sex, residence, wealth, and survey wave. The authors report that GPT-4o-mini achieves balanced accuracy of 58% versus the random forest's 57%, with substantially higher sensitivity (77.5% vs. 23.4%) and lower specificity (38.6% vs. 90.7%), and that AUROC is 0.632 versus 0.685. Fairness results show relative parity by sex but marked disparities by residence and wealth; temporal results are described as stable across waves.

Significance. If the central claim were fully established, the result would be noteworthy: a general-purpose LLM, without task-specific training, reaching balanced-accuracy parity with a supervised baseline on tabular public-health survey data would be practically interesting for rapid screening and for discussions about foundation models in public health. The paper also contributes a useful descriptive comparison across fairness and temporal strata, and the robustness check that removes the wealth index from the prompt is a thoughtful attempt to separate explicit from implicit socioeconomic signals. However, the headline 'zero-shot' claim is not yet supported because prompt construction is not shown to be independent of the evaluation data, and no contamination check is reported. The paper's descriptive statistics are mostly internally consistent, but the temporal-table inconsistency and threshold-dependent comparison need to be addressed before the main conclusion is credible.

major comments (4)

[Section 2.2] The prompt template 'was iteratively refined through empirical validation,' but no development/evaluation split is reported. Because the final model is then evaluated on all 17,106 records, the reported balanced accuracy may reflect prompt selection on the test set. Prompt tuning is a form of model selection, and without a held-out split the phrase 'zero-shot' is misleading. Please report how many prompt variants were tried, what criterion was used, and either fix a prompt a priori or evaluate on a hold-out subset (e.g., the most recent BDHS wave) that was not used during prompt development.
[Section 2.2 / Section 3] No contamination check is performed. GPT-4o-mini's pretraining data are not described, and DHS-derived statistics, reports, and related public-health text are widely available online. The model's high sensitivity (77.5%) and very high sensitivity for rural and poorest groups could partly reflect memorized population-level associations rather than reasoning from the prompt. A concrete test would be to query the model for BDHS stunting statistics by year and wealth category and compare its answers with the actual data, or to evaluate on a newly collected/private wave. Until then, the zero-shot inference claim is not fully established.
[Table 4 / Section 3] The temporal-robustness rows are internally inconsistent: the listed n for BDHS rounds are 219 + 2587 + 1661 + 1676 + 909 = 7052, not the stated overall n of 17,106. The paper claims 'stable predictive behaviour across BDHS waves' on the full dataset, but the table appears to cover only a subset. Please reconcile the total or explain which children are included in the temporal analysis, and clarify why the 2007 round has n = 219. This inconsistency affects the reliability of the temporal-robustness conclusion.
[Table 4 / Section 3] The 'comparable balanced accuracy' claim is based on a single operating point (threshold at 0.5). GPT-4o-mini's AUROC (0.632) is below the random forest's (0.685), and the two models operate at very different sensitivity/specificity trade-offs (77.5/38.6 vs. 23.4/90.7). At a different threshold, the comparison could reverse or widen. Also, the random forest results are reported from 5-fold CV without error bars or confidence intervals, and the LLM is evaluated on the full sample. Please report AUROC with confidence intervals, threshold-independent metrics (e.g., PR-AUC), and, ideally, multiple operating points or calibration curves.

minor comments (5)

[Section 2.2] Random forest hyperparameters are only partially specified (class_weight='balanced', random_state=42). For reproducibility, please report the number of trees, maximum depth, and any other tuned hyperparameters, or state that defaults were used.
[Table numbering] Table 3 (the prompt example) appears after Table 1, but there is no Table 2 in the manuscript. Please renumber the tables sequentially.
[Section 3] The fairness and temporal subgroups are reported without confidence intervals. Given the varying subgroup sizes (e.g., 219 in 2007, 3,568 in poorest wealth), confidence intervals would help assess whether observed differences are meaningful.
[Section 3] The discussion of temporal changes says 'This shift may reflect temporal changes in population characteristics and healthcare access patterns,' but no supporting analysis links the sensitivity/specificity shift to specific covariate changes. A brief examination or a more cautious statement would strengthen the claim.
[Section 2.2] The paper states 'GPT-4o-mini was used as a pretrained LLM' but does not specify the exact model version or API snapshot. Since model updates can change behavior, please include this information.

Circularity Check

0 steps flagged

Empirical zero-shot evaluation with no derivation chain; no circularity found.

full rationale

The paper is a direct empirical evaluation rather than a derivation: child records are serialized into prompts, GPT-4o-mini produces token-level probabilities, and those predictions are compared with observed stunting status defined by HAZ < -2. There are no fitted equations, no parameter estimates are renamed as predictions, and no load-bearing self-citations appear in the reference list. The main evaluation threat is in §2.2, where 'The prompt template was iteratively refined through empirical validation' without a reported held-out prompt-development split; if the same 17,106 records guided prompt selection, the zero-shot result would be contaminated. However, the paper does not state that the evaluation records were used for refinement, and pretraining contamination is an external validity risk rather than an in-paper reduction of the result to its inputs. Similarly, the Table 4 temporal rows sum to 7,052 rather than 17,106, which indicates incomplete documentation but not circular reasoning. No quote from the paper exhibits a prediction that is equivalent by construction to an input or to a self-citation, so no circular step is established.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim is empirical and rests on standard ML evaluation assumptions plus an unverified cleanliness assumption about the pretrained LLM. No new entities are postulated.

free parameters (2)

Prompt template (iteratively refined) = not specified
Section 2.2 says the prompt template was iteratively refined through empirical validation; if tuned on the same data, this is a hand-optimized component of the 'zero-shot' pipeline and may leak information.
Random forest hyperparameters = class_weight='balanced', random_state=42
Baseline model settings chosen without reported tuning; affects the supervised comparison.

axioms (4)

domain assumption BDHS data are representative and anthropometric measurements are accurate.
Section 2.1 uses BDHS data; stunting labels depend on height-for-age z-scores from WHO growth standards.
domain assumption GPT-4o-mini pretraining does not contain the evaluation data.
Zero-shot inference in Section 2.2 assumes no memorization of BDHS/DHS-derived text; not tested.
domain assumption Serialized list prompts preserve the information needed for classification.
Section 2.2 relies on TabLLM-style list serialization to convey tabular semantics.
domain assumption Children are independent observations despite survey clustering or stratification.
The analysis treats each child as i.i.d.; no survey-design or clustering adjustments are described.

pith-pipeline@v1.3.0-daily-deepseek · 6186 in / 9522 out tokens · 97121 ms · 2026-08-03T13:57:38.097092+00:00 · methodology

0 comments

read the original abstract

Child malnutrition remains a major public health challenge in low- and middle-income countries, particularly in South Asia, where early identification of vulnerable children is critical for timely intervention and resource allocation. This study aims to evaluate the feasibility, fairness, and temporal robustness of using a pretrained large language model (LLM) in a zero-shot setting for child stunting prediction using population health survey data. Using Bangladesh Demographic and Health Survey (BDHS) data collected between 2007 and 2022, we transformed maternal, child, healthcare, and household characteristics into semantically interpretable prompt-based representations and evaluated GPT-4o-mini for zero-shot stunting prediction, comparing its performance against a random forest baseline and assessing fairness across demographic and socioeconomic groups as well as temporal robustness across survey waves. The results demonstrate that zero-shot inference using GPT-4o-mini achieved comparable balanced accuracy to the supervised baseline while exhibiting substantially higher sensitivity for identifying stunting cases, relatively consistent performance across child sex groups, and stable predictive behaviour across BDHS waves; however, important fairness disparities were observed across residence and household wealth categories, highlighting the need for further investigation before deployment of foundation models in public health prediction settings.

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 1 linked inside Pith

[1]

C. G. Victora, L. Adair, C. Fall, P. C. Hallal, R. Martorell, L. Richter, H. S. Sachdev, Maternal and child undernutrition: consequences for adult health and human capital, The Lancet 371 (2008) 340–357. doi:10.1016/S0140-6736(07)61692-4

doi:10.1016/s0140-6736(07)61692-4 2008
[2]

R. E. Black, C. G. Victora, S. P. Walker, Z. A. Bhutta, P. Christian, M. de Onis, M. Ezzati, S. Grantham- McGregor, J. Katz, R. Martorell, R. Uauy, Maternal and child undernutrition and overweight in low-income and middle-income countries, The Lancet 382 (2013) 427–451. doi: 10.1016/ S0140-6736(13)60937-X

2013
[3]

UNICEF, World Health Organization, World Bank Group, Levels and trends in child malnutrition: Unicef/who/world bank group joint child malnutrition estimates: Key findings of the 2023 edition,

2023
[4]

URL: https://dhsprogram.com/publications/ publication-fr386-dhs-final-reports.cfm

National Institute of Population Research and Training (NIPORT), ICF, Bangladesh demo- graphic and health survey 2022: Final report, 2024. URL: https://dhsprogram.com/publications/ publication-fr386-dhs-final-reports.cfm

2022
[5]

Talukder, B

A. Talukder, B. Ahammed, Machine learning algorithms for predicting malnutrition among under- five children in bangladesh, Nutrition 78 (2020) 110861. doi:10.1016/j.nut.2020.110861

arXiv 2020
[6]

J. R. Khan, J. H. Tomal, E. Raheem, Model and variable selection using machine learning methods with applications to childhood stunting in bangladesh, Informatics for Health and Social Care 46 (2021) 425–442. doi:10.1080/17538157.2021.1904938

arXiv 2021
[7]

M. M. Islam, N. M. S. J. Kibria, S. Kumar, D. C. Roy, M. R. Karim, Prediction of undernutrition and identification of its influencing predictors among under-five children in bangladesh using explainable machine learning algorithms, PLOS ONE 19 (2024) e0315393. doi: 10.1371/journal. pone.0315393

doi:10.1371/journal 2024
[8]

Bommasani, D

R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill, E. Brynjolfsson, S. Buch, D. Card, R. Castellon, N. Chatterji, A. Chen, K. Creel, J. Q. Davis, D. Demszky, C. Donahue, M. Doumbouya, E. Durmus, S. Ermon, J. Etchemendy, K. Ethayarajh, L. Fei-Fei, C. Finn, T. Gale, L. Gillespie, K. Go...

Pith/arXiv arXiv doi:10.48550/arxiv.2108.07258 2021
[9]

Singhal, S

K. Singhal, S. Azizi, T. Tu, S. S. Mahdavi, J. Wei, H. W. Chung, N. Scales, A. Tanwani, H. Cole- Lewis, S. Pfohl, P. Payne, M. Seneviratne, P. Gamble, C. Kelly, A. Babiker, N. Schärli, A. Chowd- hery, P. Mansfield, D. Demner-Fushman, B. A. y Arcas, D. Webster, G. S. Corrado, Y. Matias, K. Chou, J. Gottweis, N. Tomasev, Y. Liu, A. Rajkomar, J. Barral, C. S...

doi:10.1038/s41586-023-0 2023
[10]

A. J. Thirunavukarasu, D. S. J. Ting, K. Elangovan, L. Gutierrez, T. Y. Tan, D. S. W. Ting, Large language models in medicine, Nature Medicine 29 (2023) 1930–1940. doi: 10.1038/ s41591-023-02448-8

2023
[11]

URL: https:// dhsprogram.com

ICF International, The DHS program – demographic and health surveys, 2024. URL: https:// dhsprogram.com

2024
[12]

URL: https://www.who

World Health Organization, WHO child growth standards: Length/height-for-age, weight-for-age, weight-for-length, weight-for-height and body mass index-for-age, 2006. URL: https://www.who. int/tools/child-growth-standards

2006
[13]

Hegselmann, A

S. Hegselmann, A. Buendia, H. Lang, M. Agrawal, X. Jiang, D. Sontag, TabLLM: Few-shot clas- sification of tabular data with large language models, in: F. Ruiz, J. Dy, J.-W. van de Meent (Eds.), Proceedings of The 26th International Conference on Artificial Intelligence and Statis- tics, volume 206 ofProceedings of Machine Learning Research, PMLR, 2023, pp...

2023
[2023]

URL: https://www.who.int/publications/i/item/9789240073791

arXiv

[1] [1]

C. G. Victora, L. Adair, C. Fall, P. C. Hallal, R. Martorell, L. Richter, H. S. Sachdev, Maternal and child undernutrition: consequences for adult health and human capital, The Lancet 371 (2008) 340–357. doi:10.1016/S0140-6736(07)61692-4

doi:10.1016/s0140-6736(07)61692-4 2008

[2] [2]

R. E. Black, C. G. Victora, S. P. Walker, Z. A. Bhutta, P. Christian, M. de Onis, M. Ezzati, S. Grantham- McGregor, J. Katz, R. Martorell, R. Uauy, Maternal and child undernutrition and overweight in low-income and middle-income countries, The Lancet 382 (2013) 427–451. doi: 10.1016/ S0140-6736(13)60937-X

2013

[3] [3]

UNICEF, World Health Organization, World Bank Group, Levels and trends in child malnutrition: Unicef/who/world bank group joint child malnutrition estimates: Key findings of the 2023 edition,

2023

[4] [4]

URL: https://dhsprogram.com/publications/ publication-fr386-dhs-final-reports.cfm

National Institute of Population Research and Training (NIPORT), ICF, Bangladesh demo- graphic and health survey 2022: Final report, 2024. URL: https://dhsprogram.com/publications/ publication-fr386-dhs-final-reports.cfm

2022

[5] [5]

Talukder, B

A. Talukder, B. Ahammed, Machine learning algorithms for predicting malnutrition among under- five children in bangladesh, Nutrition 78 (2020) 110861. doi:10.1016/j.nut.2020.110861

arXiv 2020

[6] [6]

J. R. Khan, J. H. Tomal, E. Raheem, Model and variable selection using machine learning methods with applications to childhood stunting in bangladesh, Informatics for Health and Social Care 46 (2021) 425–442. doi:10.1080/17538157.2021.1904938

arXiv 2021

[7] [7]

M. M. Islam, N. M. S. J. Kibria, S. Kumar, D. C. Roy, M. R. Karim, Prediction of undernutrition and identification of its influencing predictors among under-five children in bangladesh using explainable machine learning algorithms, PLOS ONE 19 (2024) e0315393. doi: 10.1371/journal. pone.0315393

doi:10.1371/journal 2024

[8] [8]

Bommasani, D

R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill, E. Brynjolfsson, S. Buch, D. Card, R. Castellon, N. Chatterji, A. Chen, K. Creel, J. Q. Davis, D. Demszky, C. Donahue, M. Doumbouya, E. Durmus, S. Ermon, J. Etchemendy, K. Ethayarajh, L. Fei-Fei, C. Finn, T. Gale, L. Gillespie, K. Go...

Pith/arXiv arXiv doi:10.48550/arxiv.2108.07258 2021

[9] [9]

Singhal, S

K. Singhal, S. Azizi, T. Tu, S. S. Mahdavi, J. Wei, H. W. Chung, N. Scales, A. Tanwani, H. Cole- Lewis, S. Pfohl, P. Payne, M. Seneviratne, P. Gamble, C. Kelly, A. Babiker, N. Schärli, A. Chowd- hery, P. Mansfield, D. Demner-Fushman, B. A. y Arcas, D. Webster, G. S. Corrado, Y. Matias, K. Chou, J. Gottweis, N. Tomasev, Y. Liu, A. Rajkomar, J. Barral, C. S...

doi:10.1038/s41586-023-0 2023

[10] [10]

A. J. Thirunavukarasu, D. S. J. Ting, K. Elangovan, L. Gutierrez, T. Y. Tan, D. S. W. Ting, Large language models in medicine, Nature Medicine 29 (2023) 1930–1940. doi: 10.1038/ s41591-023-02448-8

2023

[11] [11]

URL: https:// dhsprogram.com

ICF International, The DHS program – demographic and health surveys, 2024. URL: https:// dhsprogram.com

2024

[12] [12]

URL: https://www.who

World Health Organization, WHO child growth standards: Length/height-for-age, weight-for-age, weight-for-length, weight-for-height and body mass index-for-age, 2006. URL: https://www.who. int/tools/child-growth-standards

2006

[13] [13]

Hegselmann, A

S. Hegselmann, A. Buendia, H. Lang, M. Agrawal, X. Jiang, D. Sontag, TabLLM: Few-shot clas- sification of tabular data with large language models, in: F. Ruiz, J. Dy, J.-W. van de Meent (Eds.), Proceedings of The 26th International Conference on Artificial Intelligence and Statis- tics, volume 206 ofProceedings of Machine Learning Research, PMLR, 2023, pp...

2023

[14] [2023]

URL: https://www.who.int/publications/i/item/9789240073791

arXiv