Performance Analysis of Post-Training Quantization for CNN-based Conjunctival Pallor Anemia Detection

Sebastian A. Cruz Romero; Wilfredo E. Lugo Beauchamp

REVIEW 4 major objections 5 minor 16 references

Performance Analysis of Post-Training Quantization for CNN-based Conjunctival Pallor Anemia Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a MobileNetV2 CNN detects anemia from conjunctival pallor with 93.13% accuracy and F1 0.9428, and that FP16 post-training quantization preserves 92.50% accuracy and F1 0.9377 while INT8 and INT4 degrade sharply.

desk verdict Plausible qualitative result, but the paper's mixed-up validation/test numbers and missing 5-fold evidence make the headline claim unsupported as written. read the letter →

arxiv 2507.15151 v1 pith:NUFM6QBN submitted 2025-07-20 eess.IV cs.AIcs.CV

Sebastian A. Cruz Romero , Wilfredo E. Lugo Beauchamp This is my paper

classification eess.IVcs.AIcs.CV

keywords conjunctivalpalloranemiadetectionMobileNetpost-trainingquantizationFP16INT8edgedeploymentcomputer-aideddiagnosis

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Anemia is usually confirmed by a blood test, which is hard to obtain in low-resource clinics. This paper claims that a lightweight convolutional network (MobileNetV2) trained on photographs of children's conjunctival pallor, the paleness of the lower eyelid, can classify anemia with 93.13% accuracy and an F1 score of 0.9428 on the CP-AnemiC dataset. It then applies post-training quantization, compressing the trained model to lower-precision arithmetic, and reports that FP16 keeps the model essentially intact, with 92.50% accuracy and F1 0.9377, while INT8 and INT4 cause severe drops. If correct, this means a small FP16 model could power offline anemia screening on mobile hardware, and 8-bit or 4-bit compression is not usable in this pipeline without further work.

What carries the argument

The load-bearing object is a fine-tuned MobileNetV2 convolutional network, converted to an exchangeable graph format and subjected to post-training quantization at FP32, FP16, INT8, and INT4. Quantization uses the uniform mapping $q_x = \mathrm{round}(x/s_x)$ with scale $s_x = \max(|x|)/(2^{b-1}-1)$: weights are scaled per channel, activations per tensor, and the scale is set from the maximum absolute value on a calibration set. FP16 conversion rounds values to reduced mantissa precision, while INT4 uses block-wise weight quantization with activations left unquantized. The comparison then turns on how each bit-width changes accuracy, precision, recall, F1, AUC, model size, and execution time.

What would settle it

Track participant identities across the five folds in the dataset metadata; if any child's images appear in both training and test folds, recompute the FP32 and FP16 metrics on strictly patient-disjoint splits. If 93.13% and 92.50% accuracy do not survive, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a MobileNetV2 convolutional network, fine-tuned end-to-end on the CP-AnemiC dataset, achieves state-of-the-art anemia detection from conjunctival pallor, and that post-training FP16 quantization preserves nearly all of that performance: accuracy 92.50% versus 93.13%, F1 0.9377 versus 0.9428, and AUC 0.9654 versus 0.9657. The same comparison shows INT8 falling to 71.25% accuracy and F1 0.7607, and INT4 collapsing to 43.13% accuracy with F1 0.0196. The authors interpret this as evidence that lightweight 16-bit models are viable for edge deployment in low-resource settings, while aggressive integer quantization is not yet viable. The reported validation table also records a best F1 of 0.9773, while the headline inference numbers used for comparison are the FP32 and FP16 values.

Load-bearing premise

The central claim assumes that the 5-fold cross-validation splits are patient-disjoint, meaning no child appears in both training and test, and that the fold numbers in the validation table label independent test folds, but the paper never states this.

Editorial extensions

If this is right

A 4.61 MB FP16 model with about 37 ms execution time is small and fast enough to be a plausible component of an offline mobile screening tool.
The near-identical FP32 and FP16 metrics mean that, for this task, the precision bottleneck is not 16-bit arithmetic but the aggressive integer representations tested.
The INT8 and INT4 results, as configured, are below the accuracy and F1 levels needed for diagnostic use, so they are not deployment candidates in this pipeline.
Because INT8 and INT4 are larger and slower than FP16 in this setup, quantization gains cannot be assumed from bit-width alone and must be measured on the target deployment stack.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If a patient-disjoint re-run reproduces these numbers, the FP16 result should transfer to new children and clinics far better than a non-disjoint split would suggest; the paper does not state the split rule, so this is the first thing to verify.
The INT4 recall near zero suggests that quantizing only weights while leaving activations in floating point biases the model toward the non-anemic class, so mixed-precision activation quantization is a natural next experiment.
The non-monotonic model sizes imply that quantization metadata and dequantization overhead can exceed the raw savings of low bit-widths; an integer-only runtime without those overheads could recover the expected speed and memory gains.
Because FP16 and FP32 are nearly identical, further accuracy gains are more likely to come from better data and labels than from numerical precision.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a small, straightforward PTQ comparison on a public dataset, and the qualitative finding—FP16 preserves accuracy, INT8/INT4 don't—is probably correct. But the paper's reporting is messy enough that the headline numbers can't be trusted as-is.

What's new and good: applying MobileNetV2 to CP-AnemiC with FP16/INT8/INT4 post-training quantization is a legitimate extension, and I don't think that specific comparison is in the prior literature. The paper also gives latency and model-size numbers, which is more than many such studies do. The trend in Table IV (FP32≈FP16, INT8 drops, INT4 collapses) is visible and consistent with general quantization experience. That's a useful data point for anyone building a mobile anemia screener.

The soft spots are real. The abstract reports F1=0.9773, which the text later says was a validation snapshot at epoch 26; the conclusion reports F1=0.9428 for the same FP32 model, but the paper never explains which is which or how the test-fold aggregates were computed. Tables II and III are labeled 'Fold' but the row labels (74, 98, 100... and 12, 13, 11...) look like training epochs, not fold indices. If those are epoch logs, then the claimed 5-fold CV is not actually shown. There's also no statement about whether patients are unique across images, no per-fold metrics, and no code or artifacts. The model-size table shows INT4 (17.75 MB) larger than FP32 (9.13 MB), which is counterintuitive and needs an explanation. Equation (2) has a typo (tp/(tp+tp)). And 'state-of-the-art' is asserted without comparing to the CP-AnemiC benchmark [7], which is the obvious baseline.

None of this makes me think the underlying experiment is fake; the qualitative pattern is too mundane to fabricate. But as written, the central claim isn't reproducible, and the inconsistencies would waste a reader's time.

My recommendation: don't desk-reject, but send it to a technical reviewer with clear instructions to focus on the evaluation protocol. If the authors can supply per-fold test metrics, patient-level split guarantees, and a proper comparison to [7], this could be a solid short paper. If not, it should be rejected.

Referee Report

4 major / 5 minor

Summary. The paper evaluates a MobileNet-based CNN for anemia detection from conjunctival pallor images using the CP-AnemiC dataset, and analyzes the effect of post-training quantization at FP32, FP16, INT8, and INT4 precision. The authors report that the unquantized model achieves 93.13% accuracy and an F1 score of 0.9428, that FP16 preserves performance (92.50% accuracy, F1 0.9377), and that INT8 and INT4 degrade substantially. The paper claims state-of-the-art performance on CP-AnemiC and motivates further study of quantization for mobile healthcare deployment.

Significance. If the reported evaluation is sound, the paper provides a useful empirical data point on the quantization-accuracy trade-off for conjunctival pallor anemia screening, a task of practical interest for low-resource settings. The qualitative trend—FP16 near FP32, INT8 worse, INT4 collapse—is visible in Table IV and is the kind of result that can inform deployment decisions. However, the manuscript as written does not allow a reader to verify the headline metrics, because the cross-validation protocol is not traceable from the reported tables and the abstract misreports the unquantized F1 score as 0.9773. These issues currently block acceptance.

major comments (4)

[II-B, III-A, Tables II/III/VI] The 5-fold cross-validation described in Section II-B is not reflected in the reported tables. The 'Fold' columns in Table II (rows labeled 74, 98, 100, ...) and Table III (rows labeled 12, 13, 11, ...) contain numbers that look like training epochs, not fold indices; Section III-A states that the best validation F1 of 0.9773 was observed at epoch 26, confirming that Table III is a validation-by-epoch table. No per-fold test-set metrics, fold assignments, or patient-exclusivity guarantees are provided. Consequently, the headline FP32 numbers in Table VI (accuracy 0.9313, F1 0.9428) cannot be verified as aggregated held-out test results, and the central claim that MobileNetV2 with FP16 PTQ is viable for mobile screening is unsupported as reported.
[Abstract, III-A, IV] The abstract reports an F1 score of 0.9773 for the model, but this number matches the validation snapshot in Table III (row '13'), not the unquantized test F1 of 0.9428 reported in Table VI and the conclusion. The abstract should report the same held-out performance as the conclusion, or clearly label the 0.9773 as a validation-snapshot number. Presenting the validation F1 as the model's performance is misleading and must be corrected.
[II-B, Eq. (2)] Equation (2) defines precision as tp/(tp+tp), which equals 1/2 for any tp>0 and is not a valid precision measure. The correct definition is tp/(tp+fp). While the numerical results in Tables II–IV presumably use the correct formula, the error in the methodology section affects the paper's technical soundness and must be fixed.
[IV] The claim of 'state-of-the-art performance on the CP-AnemiC dataset' is not supported by any comparison with the CP-AnemiC benchmark (reference [7]) or with prior models on the same dataset. A quantitative comparison table against existing results is required to substantiate this claim.

minor comments (5)

[II-B, II-C, Fig. 2] The architecture is referred to as 'MobileNet' in the abstract, Section II-B, and Figure 2, but Section II-C and the conclusion refer to 'MobileNetV2'. The authors should state which architecture is used and cite the corresponding paper, as this is essential for reproducibility.
[III-B, III-D] Sections III-B and III-D contain duplicated text on memory and execution time; the paragraph beginning 'Table VI demonstrate trade-offs' appears verbatim in both sections. One of the duplicates should be removed.
[Eq. (5)] The AUC definition is written with a nonstandard integral notation; the standard form is AUC = ∫ TPR d(FPR) evaluated over the ROC curve with appropriate limits.
[Table VI] The execution-time measurements (e.g., 48.6 ms ± 235 µs) do not report the hardware, batch size, or inference framework used. This information should be added for reproducibility.
[III-C] The phrase 'layered a top of existing architecture' should read 'layered on top of the existing architecture'; there are also minor grammatical issues in Table IV captions and Section III-B.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports empirical measurements, with standard PTQ equations that do not embed the target accuracy results.

full rationale

The paper's derivation chain is essentially an empirical evaluation pipeline: MobileNetV2 is fine-tuned on the public CP-AnemiC dataset, then quantized with NVIDIA ModelOpt/TensorRT, and metrics are measured. The quantization formulas (FP16 Value = Round(x/2^k)*2^k; qx = Round(x/sx), sx = max(|x|)/(2^{b-1}-1)) are standard definitions and do not presuppose the reported accuracies. No fitted parameter is renamed as a prediction, and the authors do not rely on self-citations or imported uniqueness theorems to force their conclusions. The unsupported 'state-of-the-art' claim and the inconsistency between the abstract's validation F1=0.9773 and the inference F1=0.9428 are validity or reporting concerns, but they are not cases where a claimed derivation reduces to its own inputs by construction. Accordingly, no circular step is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on dataset quality, transfer learning, patient-disjoint splitting, and tooling fidelity. No new physical entities or mechanisms are introduced. The main fitted quantities are standard training hyperparameters and PTQ calibration scale factors, which do not add scientific free parameters beyond normal ML practice.

free parameters (6)

Learning rate = 1e-4
Hand-chosen training hyperparameter in Section II-B; standard value, not fitted to the target metric.
Batch size = 32
Hand-chosen in Section II-B; standard training choice.
Early stopping patience = 10 epochs
Hand-chosen in Section II-B; defines which checkpoint is selected and therefore affects reported metrics.
Max training epochs = 150
Upper bound in Section II-B; early stopping usually stops earlier.
AWQ block size = 128 elements
Chosen quantization configuration for INT4 in Section II-C; affects quantization error and model size.
Quantization scale factors = Per-tensor for activations, per-channel for weights in INT8; per-block for INT4
Computed from calibration data by MaxCalibrator; the quantized model's behavior depends on these fitted constants, but they are standard PTQ calibration parameters rather than scientific model parameters.

assumptions (5)

domain assumption Conjunctival pallor is a valid diagnostic indicator for anemia
Section I cites clinical literature [5] for the relationship between pallor and hemoglobin; the prediction task relies on this.
domain assumption CP-AnemiC labels are correct and the WHO threshold of hemoglobin <11 g/dL defines anemia
Section II-A uses the dataset's annotation; label errors would directly propagate into the reported metrics.
domain assumption ImageNet-pretrained features transfer to conjunctival images
Section II-B fine-tunes ImageNet-pretrained MobileNet; if transfer failed, the reported accuracy would not hold.
domain assumption The 5-fold cross-validation splits are patient-disjoint
Section II-B describes 5-fold CV but does not state whether images from the same child are grouped; leakage would inflate metrics.
domain assumption TensorRT and ModelOpt faithfully implement the described quantization
Section II-C relies on NVIDIA ModelOpt with TensorRT backend; the measured model sizes and latencies depend on tooling versions and configurations.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Performance Analysis of Post-Training Quantization for CNN-based Conjunctival Pallor Anemia Detection." pith.science (2026). https://pith.science/paper/NUFM6QBN

@misc{pith2026250715151,
  author       = {Pith},
  title        = {Pith review of: Performance Analysis of Post-Training Quantization for CNN-based Conjunctival Pallor Anemia Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NUFM6QBN}},
  note         = {Machine review of arXiv:2507.15151}
}

read the original abstract

Anemia is a widespread global health issue, particularly among young children in low-resource settings. Traditional methods for anemia detection often require expensive equipment and expert knowledge, creating barriers to early and accurate diagnosis. To address these challenges, we explore the use of deep learning models for detecting anemia through conjunctival pallor, focusing on the CP-AnemiC dataset, which includes 710 images from children aged 6-59 months. The dataset is annotated with hemoglobin levels, gender, age and other demographic data, enabling the development of machine learning models for accurate anemia detection. We use the MobileNet architecture as a backbone, known for its efficiency in mobile and embedded vision applications, and fine-tune our model end-to-end using data augmentation techniques and a cross-validation strategy. Our model implementation achieved an accuracy of 0.9313, a precision of 0.9374, and an F1 score of 0.9773 demonstrating strong performance on the dataset. To optimize the model for deployment on edge devices, we performed post-training quantization, evaluating the impact of different bit-widths (FP32, FP16, INT8, and INT4) on model performance. Preliminary results suggest that while FP16 quantization maintains high accuracy (0.9250), precision (0.9370), and F1 Score (0.9377), more aggressive quantization (INT8 and INT4) leads to significant performance degradation. Overall, our study supports further exploration of quantization schemes and hardware optimizations to assess trade-offs between model size, inference time, and diagnostic accuracy in mobile healthcare applications.

Figures

Figures reproduced from arXiv: 2507.15151 by the authors.

**Figure 1.** Sample images of conjunctival pallor with an example region-of-interest from the CP-AnemiC dataset. The first row represents images from the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** MobileNet architecture used as a backbone for our anemia detection task. A conjunctival pallor region-of-interest is used as input and the model [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Training and Validation Performance of our fine-tuned MobileNet for Anemia Detection after 26 epochs. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 11 canonical work pages

[7]

Cp-anemic: A conjunctival pallor dataset and benchmark for anemia detection in children,

G. Dapena and et al., “Cp-anemic: A conjunctival pallor dataset and benchmark for anemia detection in children,” Pattern Recognition Letters , 2024. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2590093523000395

work page 2024
[1]

Anemia epidemiology, pathophysiology, and etiology in low- and middle-income countries,

C. M. Chaparro and P. S. Suchdev, “Anemia epidemiology, pathophysiology, and etiology in low- and middle-income countries,” Annals of the New York Academy of Sciences , vol. 1450, no. 1, pp. 15–31, 2019. [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/PMC6697587/

work page 2019
[2]

Anaemia,

World Health Organization, “Anaemia,” 2023. [Online]. Available: https://www.who.int/news-room/fact-sheets/detail/anaemia

work page 2023
[3]

Diagnosing anemia: Challenges selecting methods, addressing underlying causes, and implementing actions at the public health level,

M. N. Garcia-Casal, O. Dary, M. E. Jefferds, and S. R. Pasricha, “Diagnosing anemia: Challenges selecting methods, addressing underlying causes, and implementing actions at the public health level,” Annals of the New York Academy of Sciences, vol. 1524, no. 1, pp. 37–50, 2023. [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/PMC10880862/

work page 2023
[4]

Emerging point-of-care technologies for anemia detection,

R. An, Y . Huang, Y . Man, R. W. Valentine, E. Kucukal, U. Goreke, Z. Sekyonda, C. Piccone, A. Owusu-Ansah, S. Ahuja, J. A. Little, and U. A. Gurkan, “Emerging point-of-care technologies for anemia detection,” Lab on a Chip , vol. 21, no. 10, pp. 1843–1865, 2021. [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/PMC8875318/

work page 2021
[5]

The relation of conjunctival pallor to the presence of anemia,

T. N. Sheth, N. K. Choudhry, M. Bowes, and A. S. Detsky, “The relation of conjunctival pallor to the presence of anemia,” Journal of General Internal Medicine , vol. 12, no. 2, pp. 102–106, 1997

work page 1997
[6]

An unacceptably high burden of anaemia and its predictors among young women (15–24 years) in low and middle income countries; setback to sdg progress,

M. W. Merid, D. Chilot, and A. Z. Alem, “An unacceptably high burden of anaemia and its predictors among young women (15–24 years) in low and middle income countries; setback to sdg progress,” BMC Public Health , vol. 23, no. 1292, 2023. [Online]. Available: https://doi.org/10.1186/s12889-023-16187-5

work page doi:10.1186/s12889-023-16187-5 2023
[8]

Internet of intelligent things: A convergence of embedded systems, edge computing and machine learning,

R. Sharma, A. Alharbi, A. Alsarhan et al. , “Internet of intelligent things: A convergence of embedded systems, edge computing and machine learning,” Journal of King Saud University - Computer and Information Sciences , 2024. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2542660524000945

work page 2024

Show all 16 references

[10]

A white paper on neural network quantization,

M. Nagel, M. Fournarakis, R. A. Amjad, Y . Bondarenko, M. van Baalen, and T. Blankevoort, “A white paper on neural network quantization,” CoRR, vol. abs/2106.08295, 2021. [Online]. Available: https://arxiv.org/abs/2106.08295

2021 arXiv
[11]

A survey of quantization methods for efficient neural network inference,

A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, and K. Keutzer, “A survey of quantization methods for efficient neural network inference,” 2021. [Online]. Available: https://arxiv.org/abs/2103.13630

2021 arXiv
[12]

Mobilenets: Efficient convolutional neural networks for mobile vision applications,

A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient convolutional neural networks for mobile vision applications,” 2017. [Online]. Available: https://arxiv.org/abs/1704.04861

2017 arXiv
[13]

Imagenet: A large-scale hierarchical image database,

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition . Ieee, 2009, pp. 248–255

2009
[14]

Onnx runtime,

O. R. developers, “Onnx runtime,” https://onnxruntime.ai/, 2021, ver- sion: x.y.z

2021
[15]

[Online]

NVIDIA, TensorRT Model Optimizer: PyTorch Quantization Guide, 2024, accessed: 2024-11-27. [Online]. Available: https://nvidia.github.io/TensorRT-Model-Optimizer/index.html

2024
[16]

[Online]

NVIDIA Corporation, TensorRT Developer Guide , 2024, accessed: 2024-11-23. [Online]. Available: https://docs.nvidia.com/deeplearning/tensorrt/pdf/TensorRT-Developer- Guide.pdf

2024
[2016]

Available: http://arxiv.org/abs/1605.07678

[Online]. Available: http://arxiv.org/abs/1605.07678

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [7]

Cp-anemic: A conjunctival pallor dataset and benchmark for anemia detection in children,

G. Dapena and et al., “Cp-anemic: A conjunctival pallor dataset and benchmark for anemia detection in children,” Pattern Recognition Letters , 2024. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2590093523000395

work page 2024

[2] [1]

Anemia epidemiology, pathophysiology, and etiology in low- and middle-income countries,

C. M. Chaparro and P. S. Suchdev, “Anemia epidemiology, pathophysiology, and etiology in low- and middle-income countries,” Annals of the New York Academy of Sciences , vol. 1450, no. 1, pp. 15–31, 2019. [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/PMC6697587/

work page 2019

[3] [2]

Anaemia,

World Health Organization, “Anaemia,” 2023. [Online]. Available: https://www.who.int/news-room/fact-sheets/detail/anaemia

work page 2023

[4] [3]

Diagnosing anemia: Challenges selecting methods, addressing underlying causes, and implementing actions at the public health level,

M. N. Garcia-Casal, O. Dary, M. E. Jefferds, and S. R. Pasricha, “Diagnosing anemia: Challenges selecting methods, addressing underlying causes, and implementing actions at the public health level,” Annals of the New York Academy of Sciences, vol. 1524, no. 1, pp. 37–50, 2023. [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/PMC10880862/

work page 2023

[5] [4]

Emerging point-of-care technologies for anemia detection,

R. An, Y . Huang, Y . Man, R. W. Valentine, E. Kucukal, U. Goreke, Z. Sekyonda, C. Piccone, A. Owusu-Ansah, S. Ahuja, J. A. Little, and U. A. Gurkan, “Emerging point-of-care technologies for anemia detection,” Lab on a Chip , vol. 21, no. 10, pp. 1843–1865, 2021. [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/PMC8875318/

work page 2021

[6] [5]

The relation of conjunctival pallor to the presence of anemia,

T. N. Sheth, N. K. Choudhry, M. Bowes, and A. S. Detsky, “The relation of conjunctival pallor to the presence of anemia,” Journal of General Internal Medicine , vol. 12, no. 2, pp. 102–106, 1997

work page 1997

[7] [6]

An unacceptably high burden of anaemia and its predictors among young women (15–24 years) in low and middle income countries; setback to sdg progress,

M. W. Merid, D. Chilot, and A. Z. Alem, “An unacceptably high burden of anaemia and its predictors among young women (15–24 years) in low and middle income countries; setback to sdg progress,” BMC Public Health , vol. 23, no. 1292, 2023. [Online]. Available: https://doi.org/10.1186/s12889-023-16187-5

work page doi:10.1186/s12889-023-16187-5 2023

[8] [8]

Internet of intelligent things: A convergence of embedded systems, edge computing and machine learning,

R. Sharma, A. Alharbi, A. Alsarhan et al. , “Internet of intelligent things: A convergence of embedded systems, edge computing and machine learning,” Journal of King Saud University - Computer and Information Sciences , 2024. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2542660524000945

work page 2024