An Ensemble Classification Approach in A Multi-Layered Large Language Model Framework for Disease Prediction

Ali Hamdi; Khaled Shaban; Malak Mohamed; Rokaia Emad

REVIEW 4 major objections 4 minor 30 references

An Ensemble Classification Approach in A Multi-Layered Large Language Model Framework for Disease Prediction

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A majority-voting ensemble of three Arabic BERT models, fine-tuned on LLAMA3-preprocessed patient posts, reaches 80.56% accuracy in seven-class disease prediction.

desk verdict A systematic but under-validated engineering study; the ensemble approach is plausible and the preprocessing analysis is useful, but the evaluation lacks rigor and has an internal inconsistency. read the letter →

arxiv 2509.02446 v1 pith:5YV7AZQQ submitted 2025-09-02 cs.CL cs.LG

Ali Hamdi , Malak Mohamed , Rokaia Emad , Khaled Shaban This is my paper · ORCID

classification cs.CLcs.LG

keywords ArabicmedicaltextclassificationsocialtelehealthLLM-basedpreprocessingBERTfine-tuningmajorityvotingensemblediseasepredictionLLAMA3

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

The paper proposes a multi-layer pipeline for classifying Arabic social telehealth posts into seven disease categories. Patient complaints are first passed through LLAMA3 for refinement, summarization, or named-entity extraction; each processed version is then paired with the original text to fine-tune three Arabic BERT models. A majority-voting ensemble of these models reaches a reported 80.56% accuracy, beating the best single model (75.55%) and the previous no-ensemble study. The paper also reports that ensembles built from summarized texts perform worst (63.64%), while refined-text ensembles perform best (77.12%), indicating that heavy compression loses medical signal. If the numbers hold, the framework offers a transferable recipe for Arabic medical text classification using off-the-shelf LLMs and transformer models.

What carries the argument

The load-bearing mechanism is the majority-voting ensemble over a grid of twelve base models: three Arabic BERT architectures (CAMeLBERT, AraBERT, AsafayaBERT) fine-tuned separately on four text representations (original 'post', original+refined, original+NER, original+summarized). LLAMA3 supplies the three preprocessing transformations, but the ensemble itself is a simple vote that aggregates each model's top label for a post. The diversity created by pairing original text with distinct processed views, and then crossing those views with different transformer backbones, is what the paper credits for the accuracy lift.

What would settle it

Re-run the pipeline on the same private dataset with a stratified, leakage-checked train/test split and repeated seeds; if the majority vote does not consistently exceed the best single model beyond chance, or if the gap to 80.56% collapses, the central claim is falsified. A simpler check: see whether a random 50/50 split reproduces the reported accuracies within a few points.

Watch

Extended reading notes

Core claim

The central claim is that combining LLM-based preprocessing with fine-tuned Arabic transformer models and a majority-voting ensemble improves disease classification on Arabic social telehealth data. The paper reports the best accuracy of 80.56% with this strategy, observed consistently across ensemble sizes of five, seven, eight, and ten models. The gain over the best single model (75.55% for the refined CAMeLBERT variant) is attributed to diversity among base models and text representations. The study also finds that grouping ensembles by preprocessing type favors refined texts (77.12%) over NER (75.24%) and summarized texts (63.64%), suggesting that over-summarization harms medical classif

Load-bearing premise

The 80.56% figure rests on the privately collected dataset's seven-class annotations being correct and on the train/test split being leakage-free, but the paper reports no split ratio, annotation agreement, or leakage checks.

Editorial extensions

If this is right

The 80.56% result implies a concrete recipe: refine and pair, don't replace; keeping the original text alongside processed versions outperforms using processed text alone.
Three diverse models appear sufficient to realize most ensemble gains; expanding to 12 models plateaus at the same accuracy, so practitioners can stop at a handful of diverse members.
Preprocessing choice dominates model choice in grouped ensembles: refined-text ensembles (77.12%) beat NER and post ensembles, while summarized-text ensembles (63.64%) hurt accuracy, so summary-based compression should be avoided for Arabic medical posts.
If replicated, the framework establishes a benchmark for Arabic social telehealth disease classification that future systems can compare against.

Reading between the lines

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

The reported margin between ensemble (80.56%) and best single model (75.55%) may not be statistically significant given no error bars or repeated-seed evaluation; a cross-validation study would tell whether the ensemble gain is real.
The same multi-layer pairing scheme should transfer to other low-resource languages that have a BERT-style model and access to a strong LLM, provided the base models are sufficiently diverse.
Because the dataset is private, external validation on an independent Arabic telehealth corpus is the natural next test; a drop in the ensemble's advantage there would indicate overfitting to this annotation scheme.
A direct extension: use the ensemble's disagreement signals as a confidence measure to flag posts needing human review, which the current binary vote does not do.

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

This paper is exactly what it looks like: a systematic engineering study combining LLAMA3 preprocessing with three fine-tuned Arabic BERT models and majority-vote ensembles for seven-class disease prediction on a private dataset. The best result, 80.56%, is plausible and beats the strongest single model by about five points. What's new is narrow but real: the specific pairing of LLM-generated paraphrases (refined, summarized, NER) with the original text, and then ensembling over that grid. Their prior arXiv:2502.00063 had the same models and dataset without ensembling, so this is a clear incremental extension.

The paper does a few things well. The experimental grid is systematic — twelve base models, all combinations up to size 12, plus grouping analyses by preprocessing type and by model family. The finding that refined models ensemble well while summarized models hurt is a useful practical hint. The annotation was done under medical supervision, which is better than nothing for a private dataset.

The soft spots are real and concentrated in the evaluation. The most concrete one: the paper says the 80.56% peak occurred at ensemble sizes 5, 7, 8, and 10; Figure 14's caption says 4, 6, 7, and 10. That's an internal contradiction that needs fixing. More importantly, 80.56% is the maximum over thousands of possible subsets of the twelve base models, and there is no separation between validation and test, no error bars, and no significance test. With the best single model at 75.55%, the ensemble gain might be real, but the way it is reported makes selection bias a live possibility. Also missing: the train/test split ratio, any leakage check, and inter-annotator agreement for the seven-class labels. Without those, the headline number can't be independently assessed.

The paper is honest about its scope; it doesn't overclaim a new mechanism, just a working recipe. I'd send it to peer review, but the reviewers should push for a fixed evaluation protocol, a pre-registered or at least clearly separated validation/test split, and significance testing or error bars. The internal inconsistency alone would require a revision. It's not a desk reject — the problem is topical and the negative result on summarization is useful — but it's not ready as-is.

For a reading group, it's a maybe: it's a fine example of how to report ensemble experiments poorly and well at the same time. I wouldn't cite it as a benchmark, but I'd cite it if I needed evidence about LLM preprocessing effects on Arabic clinical text.

Recommendation: engage with it, but only with the expectation of substantial revision.

Referee Report

4 major / 4 minor

Summary. This paper proposes a multi-layered framework for Arabic disease classification from social telehealth posts. LLAMA3 is used for three preprocessing variants (refinement, summarization, NER), which are paired with original text to fine-tune three Arabic BERT models (CAMeLBERT, AraBERT, AsafayaBERT), yielding twelve base models. Majority-voting ensembles over these models are evaluated on a private seven-class dataset. The paper reports a best accuracy of 80.56% and claims that ensemble strategies improve disease classification performance. It presents this as the first integration of LLM-based preprocessing, fine-tuned Arabic transformers, and ensemble learning for this task.

Significance. If the result is robust, the framework is a useful recipe for Arabic social telehealth classification, and the comparison of preprocessing variants is of practical interest. The paper does not contain a derivation, so there is no circular-fit concern; the claim is empirical. However, the evaluation is entirely internal: one private dataset, no public benchmark, no released code or data, and no fixed evaluation protocol. The paper's strengths are the systematic exploration of ensemble sizes and the direct comparison of preprocessing strategies; these are undermined by the absence of error bars, statistical tests, and a clear model-selection procedure. An honest re-analysis with a pre-registered validation/test split and significance tests would make the contribution much stronger.

major comments (4)

[Section V, Table II, Figures 3-14] The claimed 80.56% is selected as the maximum over an unstated combinatorial search of ensemble subsets. There is no separation between a model-selection validation set and a held-out test set, no error bars, and no significance tests. Moreover, the text in Section V says this accuracy was 'consistently observed across ensemble sizes of 5, 7, 8, and 10 models' while the caption of Fig. 14 says '4, 6, 7, and 10-model ensembles.' This internal inconsistency indicates the result-selection rule is not well-defined. Without a fixed protocol, the headline number may reflect selection bias rather than a genuine ensemble advantage. Please report a single pre-specified ensemble configuration (or all configurations with test-set means and standard deviations), with significance tests against the best single model (CamelBERT Refined, 75.55%).
[Section III] The dataset description is incomplete. It gives only the class distribution and states that annotation was supervised by a medical advisor. No total size, no train/validation/test split ratio, no leakage checks, and no inter-annotator agreement are reported. Since this is a private dataset, the reproducibility of the 80.56% figure depends entirely on these details. Add dataset statistics, split procedure, and if possible release the data or provide a detailed data sheet.
[Sections IV-E and V] There is a mismatch in the number of base models. Section IV-E says fine-tuning three Arabic BERT models and collecting predictions 'from all three models,' but Section V constructs ensembles from twelve base models (3 models × 4 representations). The relationship between the three-model procedure and the twelve-model ensemble is not explained. Also, the paper alternates between 'multi-class, multi-label' (Abstract, Section I) and seven disease types with accuracy as the only metric; if the annotation is single-label, the multi-label terminology should be removed; if it is multi-label, accuracy is not the appropriate evaluation metric and metrics like macro-F1 should be reported.
[Section V, Figures 3-15] The conclusion that 'ensemble strategies improve disease classification performance' is not supported by the evidence as reported. Figures 3-14 show many ensemble configurations with accuracies around 0.79-0.80, but some groups, e.g., summarized-model ensembles in Fig. 15, perform much worse (0.6364). Without a statistical comparison of the best ensemble to the best individual model, and without reporting variability across folds or seeds, the headline claim remains an observation about one selected configuration, not a general finding.

minor comments (4)

[Conclusion] Typo: 'accuracya' should be 'accuracy'.
[Table I] The Arabic text in Table I appears as garbled mojibake in the PDF; ensure the glyphs render correctly.
[Abstract] The claim 'to the best of our knowledge, this is the first work' is difficult to verify and should be softened or supported by a more thorough comparison with existing Arabic medical text classification benchmarks.
[Section II] The related work surveys general transformer/LLM results but does not position the proposed method against existing Arabic medical text classification systems; please add direct comparative baselines from the literature.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported accuracy is an empirical measurement, not a derived or fitted quantity; selection-bias concerns are a reproducibility issue, not circularity.

full rationale

This paper makes no derivation claim whose conclusion could be equivalent to its inputs. The pipeline is empirical: raw posts are passed through LLAMA3 refinement, summarization, and NER; the outputs are concatenated with the original text; three Arabic BERT models are fine-tuned; and majority voting is applied to the resulting twelve checkpoints. The headline 80.56% accuracy is a reported test-set measurement from Table II, not the output of a formula fitted to the data. Majority voting is not mathematically forced to exceed the best individual model, so the comparison in Table II has independent empirical content. The only self-citation, reference [22], is used to state that the authors previously fine-tuned the same models on the same dataset without ensembling; it motivates novelty but does not supply any load-bearing theorem, parameter, or result. There are genuine methodological concerns — the test set appears to have been used to select the best ensemble among many combinations, and Section V says 0.8056 was observed for ensemble sizes 5, 7, 8, and 10 while the Figure 14 caption says 4, 6, 7, and 10 — but these are selection-bias and reproducibility problems, not circularity. No equation, definition, or citation chain makes the reported accuracy equivalent to the paper's inputs. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim is empirical, so there are no invented entities. The framework leans on several domain assumptions: LLAMA3 preprocessing preserves task-relevant medical content, the privately annotated dataset is reliable and split without leakage, the three Arabic BERT checkpoints are suitable base models, and majority voting is appropriate. The numerical freedoms are the shared training hyperparameters and the post hoc selection of the best ensemble configuration.

free parameters (2)

Training hyperparameters = lr=1e-4, batch_size=4, epochs=25, dropout=0.05, weight_decay=0.01
Chosen by the authors and fixed for all 12 runs; changes here would likely change the reported accuracies.
Ensemble composition and size = best ensemble reaches 0.8056, selected from configs of size 2 to 12
The headline result is the maximum over many ensemble combinations explored in Section V, not a pre-registered single configuration; this selection is part of how the number was produced.

assumptions (5)

domain assumption LLAMA3 preprocessing (refinement, summarization, NER) preserves the discriminative medical information needed for seven-class disease prediction.
Section IV-B uses these outputs as the basis for all fine-tuning; no evaluation of preprocessing quality or comparison with other preprocessing is provided.
domain assumption The private dataset annotations are accurate and the train/test split is leakage-free.
Section III mentions supervision by a medical advisor but gives no split ratio, no inter-annotator agreement, and no leakage checks.
domain assumption The three Arabic transformer checkpoints provide suitable representations for Arabic medical text.
Chosen in Section IV-D from prior literature without dataset-specific selection criteria.
domain assumption Majority voting is an effective aggregation for the 12 classifiers.
Section IV-E applies equal-weight voting without analyzing error correlation, model confidence, or weighting schemes.
ad hoc to paper One shared hyperparameter configuration is adequate for all models.
Section IV-D fixes lr=1e-4, batch=4, epochs=25, dropout=0.05, weight decay=0.01 for every model; no tuning or ablation is reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An Ensemble Classification Approach in A Multi-Layered Large Language Model Framework for Disease Prediction." pith.science (2026). https://pith.science/paper/5YV7AZQQ

@misc{pith2026250902446,
  author       = {Pith},
  title        = {Pith review of: An Ensemble Classification Approach in A Multi-Layered Large Language Model Framework for Disease Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5YV7AZQQ}},
  note         = {Machine review of arXiv:2509.02446}
}

read the original abstract

Social telehealth has made remarkable progress in healthcare by allowing patients to post symptoms and participate in medical consultations remotely. Users frequently post symptoms on social media and online health platforms, creating a huge repository of medical data that can be leveraged for disease classification. Large language models (LLMs) such as LLAMA3 and GPT-3.5, along with transformer-based models like BERT, have demonstrated strong capabilities in processing complex medical text. In this study, we evaluate three Arabic medical text preprocessing methods such as summarization, refinement, and Named Entity Recognition (NER) before applying fine-tuned Arabic transformer models (CAMeLBERT, AraBERT, and AsafayaBERT). To enhance robustness, we adopt a majority voting ensemble that combines predictions from original and preprocessed text representations. This approach achieved the best classification accuracy of 80.56%, thus showing its effectiveness in leveraging various text representations and model predictions to improve the understanding of medical texts. To the best of our knowledge, this is the first work that integrates LLM-based preprocessing with fine-tuned Arabic transformer models and ensemble learning for disease classification in Arabic social telehealth data.

Figures

Figures reproduced from arXiv: 2509.02446 by the authors.

**Figure 1.** Proposed Ensembling approach in a Multi-Layered Framework for Enhancing Arabic Language Model Fine-Tuning with LLAMA3 Preprocessing [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Distribution of disease types in the dataset. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Accuracy of Top 2-Model Ensemble Combinations. The best [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 5.** Figure 5: Accuracy of Top 4-Model Ensemble Combinations. The best [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 7.** Figure 7: Accuracy of Top 6-Model Ensemble Combinations. All combinations [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]

**Figure 9.** Figure 9: Accuracy of Top 8-Model Ensemble Combinations. Matches best [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]

**Figure 11.** Figure 11: Accuracy of Top 10-Model Ensemble Combinations. Matches [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]

**Figure 13.** Figure 13: Accuracy of Top 12-Model Ensemble Combinations. The highest [PITH_FULL_IMAGE:figures/full_fig_p006_13.png]

**Figure 15.** Figure 15: Accuracy by Model Type Groups. The ensemble with refined models achieved the highest accuracy, while summarized models performed the worst. [PITH_FULL_IMAGE:figures/full_fig_p007_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

[1]

Social media use for health purposes: Systematic review,

J. Chen and Y . Wang, “Social media use for health purposes: Systematic review,” Journal of Medical Internet Research , vol. 23, no. 5, 2020

work page 2020
[2]

Mining social media data for biomedical signals and health-related behavior,

R. B. Correia, I. B. Wood, J. Bollen, and L. M. Rocha, “Mining social media data for biomedical signals and health-related behavior,” Annual Review of Biomedical Data Science , vol. 3, no. 1, pp. 433–458, 2020

work page 2020
[3]

Overview of the Sixth Social Media Mining for Health Applications (SMM4H) Shared Tasks at NAACL 2021,

A. Magge et al., “Overview of the Sixth Social Media Mining for Health Applications (SMM4H) Shared Tasks at NAACL 2021,” in Proc. of the Sixth Social Media Mining for Health Applications (SMM4H) Shared Task at NAACL, Jan. 2021

work page 2021
[4]

Evaluating large language models for health-related text classification tasks with public social media data,

Y . Guo, A. Ovadje, M. A. Al-Garadi, and A. Sarker, “Evaluating large language models for health-related text classification tasks with public social media data,” Journal of the American Medical Informatics Association, 2024

work page 2024
[5]

Text classification algorithms: A survey,

K. Kowsari, K. J. Meimandi, M. Heidarysafa, S. Mendu, L. Barnes, and D. Brown, “Text classification algorithms: A survey,” Information, vol. 10, no. 4, p. 150, Apr. 2019

work page 2019
[6]

Repeata- bility of fine-tuning large language models illustrated using QLORA,

S. S. Alahmari, L. O. Hall, P. R. Mouton, and D. B. Goldgof, “Repeata- bility of fine-tuning large language models illustrated using QLORA,” IEEE Access, 2024

work page 2024
[7]

Multi-label classification with generative large language models,

N. Niraula, S. Ayhan, B. Chidambaram, and D. Whyatt, “Multi-label classification with generative large language models,” in Proc. of the 43rd Digital Avionics Systems Conference (DASC) , 2024

work page 2024
[8]

A fine- tuned BERT-based transfer learning approach for text classification,

R. Qasim, W. H. Bangyal, M. A. Alqarni, and A. A. Almazroi, “A fine- tuned BERT-based transfer learning approach for text classification,” Journal of Healthcare Engineering , 2022

work page 2022

Show all 30 references

[9]

Nikolov-Radivchev at SemEval-2019 Task 6: Offensive tweet classification with BERT and ensembles,

V . Radivchev and A. Nikolov, “Nikolov-Radivchev at SemEval-2019 Task 6: Offensive tweet classification with BERT and ensembles,” in Proc. of the 13th Int. Workshop on Semantic Evaluation (SemEval-2019), pp. 691–695, 2019

2019
[10]

FBK @ SMM4H 2020: RoBERTa for detecting medications on Twitter,

S. Casola and A. Lavelli, “FBK @ SMM4H 2020: RoBERTa for detecting medications on Twitter,” in Proc. of the Fifth Social Media Mining for Health Applications Workshop & Shared Task , pp. 101–103, 2020

2020
[11]

Analyzing ELMo and DistilBERT on socio-political news classification,

B. B ¨uy¨uk¨oz, A. H ¨urriyeto˘glu, and A. ¨Ozg¨ur, “Analyzing ELMo and DistilBERT on socio-political news classification,” in Proc. of the Workshop on Automatic Extraction of Socio-political Events from News , pp. 9–18, USA, May 2020

2020
[12]

TorontoCL at CMCL 2021 shared task: RoBERTa with multi-stage fine-tuning for eye-tracking prediction,

B. Li and F. Rudzicz, “TorontoCL at CMCL 2021 shared task: RoBERTa with multi-stage fine-tuning for eye-tracking prediction,” in Proc. of the Workshop on Cognitive Modeling and Computational Linguistics (CMCL), pp. 4–9, 2021

2021
[13]

ZYJ123@DravidianLangTech-EACL2021: Offen- sive language identification based on XLM-RoBERTa with DPCNN,

Y . Zhao and X. Tao, “ZYJ123@DravidianLangTech-EACL2021: Offen- sive language identification based on XLM-RoBERTa with DPCNN,” in Proc. of the First Workshop on Speech and Language Technologies for Dravidian Languages, pp. 216–221, Parkville, Victoria: EACL, 2021

2021
[14]

YNU @ Dravidian-CodeMix-FIRE2020: XLM- RoBERTa for multi-language sentiment analysis,

X. Ou and H. Li, “YNU @ Dravidian-CodeMix-FIRE2020: XLM- RoBERTa for multi-language sentiment analysis,” in Proc. of the Forum for Information Retrieval Evaluation (FIRE) , pp. 4–9, 2020

2020
[15]

ASEM: Enhancing empathy in chatbot through attention-based sentiment and emotion modeling,

O. Hamad, K. Shaban, and A. Hamdi, “ASEM: Enhancing empathy in chatbot through attention-based sentiment and emotion modeling,” in Proc. of the Int. Conf. on Computational Linguistics (COLING) , pp. 1588–1601, 2024

2024
[16]

Ensemble of ELECTRA for profiling fake news spreaders,

K. A. Das, A. Baruah, F. A. Barbhuiya, and K. Dey, “Ensemble of ELECTRA for profiling fake news spreaders,” in Proc. of the Forum for Information Retrieval Evaluation (FIRE) , pp. 22–25, 2020

2020
[17]

BART-TL: Weakly-supervised topic label generation,

C. Popa and T. Rebedea, “BART-TL: Weakly-supervised topic label generation,” in Proc. of the 16th Conf. of the European Chapter of the ACL (EACL), pp. 1418–1425, 2021

2021
[18]

Using BERT and BART for query suggestion,

A. Mustar, S. Lamprier, and B. Piwowarski, “Using BERT and BART for query suggestion,” in Proc. of the CEUR Workshop on NLP , vol. 2621, 2020

2020
[19]

BART for knowledge grounded conversations,

M. de Bruyn, E. Lotfi, J. Buhmann, and W. Daelemans, “BART for knowledge grounded conversations,” in Proc. of the CEUR Workshop on NLP, vol. 2666, 2020

2020
[20]

LMRPA: Large language model-driven efficient robotic process automation for OCR,

O. H. Abdellatif, A. N. Hassan, and A. Hamdi, “LMRPA: Large language model-driven efficient robotic process automation for OCR,” in The Int. Conf. of Advanced Computing and Informatics , pp. 35–44, Springer, 2024

2024
[21]

LLM-SEM: A sentiment- based student engagement metric using LLMs for e-learning platforms,

A. Hamdi, A. A. Mazrou, and M. Shaltout, “LLM-SEM: A sentiment- based student engagement metric using LLMs for e-learning platforms,” in The Int. Conf. of Advanced Computing and Informatics , pp. 145–154, Springer, 2024

2024
[22]

A multi-layered large language model framework for disease prediction,

M. Mohamed, R. Emad, and A. Hamdi, “A multi-layered large language model framework for disease prediction,” arXiv preprint arXiv:2502.00063, 2025

2025 arXiv
[23]

BERT: Pre-training of deep bidirectional transformers for language understanding,

J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” in Proc. of the 2019 Conf. of the NAACL , pp. 4171–4186, 2019

2019
[24]

Contextual se- mantic embeddings based on transformer models for Arabic biomedical questions classification,

I. Ait Talghalit, H. Alami, and S. Ouatik El Alaoui, “Contextual se- mantic embeddings based on transformer models for Arabic biomedical questions classification,” HighTech and Innovation Journal , vol. 5, no. 4, pp. 1024, 2024

2024
[25]

StEduCov: An explored and benchmarked dataset on stance detection in tweets towards online education during COVID-19 pandemic,

O. Hamad, A. Hamdi, S. Hamdi, and K. Shaban, “StEduCov: An explored and benchmarked dataset on stance detection in tweets towards online education during COVID-19 pandemic,” Big Data and Cognitive Computing, vol. 6, no. 3, pp. 88, 2022

2022
[26]

CLASEG: Advanced multi- classification and segmentation for differential diagnosis of oral lesions using deep learning,

A. Al-Ali, A. Hamdi, M. Elshrif, et al. , “CLASEG: Advanced multi- classification and segmentation for differential diagnosis of oral lesions using deep learning,” Scientific Reports, vol. 15, no. 1, pp. 23016, 2025

2025
[27]

Arabic large language models for medical text generation,

A. Allam, S. Ahmed, A. Hamdi, and A. Mohammed, “Arabic large language models for medical text generation,” in 2025 4th Int. Conf. on Computer Technologies (ICCTech), pp. 1–6, 2025

2025
[28]

GloVe: Global vectors for word representation,

J. Pennington, R. Socher, and C. D. Manning, “GloVe: Global vectors for word representation,” in Proc. of the 2014 Conf. on Empirical Methods in NLP (EMNLP) , pp. 1532–1543, 2014

2014
[29]

AraBERT: Transformer-based model for Arabic language understanding,

W. Antoun, F. Baly, and H. Hajj, “AraBERT: Transformer-based model for Arabic language understanding,” in Proc. of the LREC 2020 Work- shop on Resources for African Indigenous Languages , pp. 9–15, 2020

2020
[30]

EDA: Easy data augmentation techniques for boosting performance on text classification tasks,

J. Wei and K. Zou, “EDA: Easy data augmentation techniques for boosting performance on text classification tasks,” Proc. of the 2019 Conf. on Empirical Methods in NLP (EMNLP) , pp. 6382–6388, 2019

2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Social media use for health purposes: Systematic review,

J. Chen and Y . Wang, “Social media use for health purposes: Systematic review,” Journal of Medical Internet Research , vol. 23, no. 5, 2020

work page 2020

[2] [2]

Mining social media data for biomedical signals and health-related behavior,

R. B. Correia, I. B. Wood, J. Bollen, and L. M. Rocha, “Mining social media data for biomedical signals and health-related behavior,” Annual Review of Biomedical Data Science , vol. 3, no. 1, pp. 433–458, 2020

work page 2020

[3] [3]

Overview of the Sixth Social Media Mining for Health Applications (SMM4H) Shared Tasks at NAACL 2021,

A. Magge et al., “Overview of the Sixth Social Media Mining for Health Applications (SMM4H) Shared Tasks at NAACL 2021,” in Proc. of the Sixth Social Media Mining for Health Applications (SMM4H) Shared Task at NAACL, Jan. 2021

work page 2021

[4] [4]

Evaluating large language models for health-related text classification tasks with public social media data,

Y . Guo, A. Ovadje, M. A. Al-Garadi, and A. Sarker, “Evaluating large language models for health-related text classification tasks with public social media data,” Journal of the American Medical Informatics Association, 2024

work page 2024

[5] [5]

Text classification algorithms: A survey,

K. Kowsari, K. J. Meimandi, M. Heidarysafa, S. Mendu, L. Barnes, and D. Brown, “Text classification algorithms: A survey,” Information, vol. 10, no. 4, p. 150, Apr. 2019

work page 2019

[6] [6]

Repeata- bility of fine-tuning large language models illustrated using QLORA,

S. S. Alahmari, L. O. Hall, P. R. Mouton, and D. B. Goldgof, “Repeata- bility of fine-tuning large language models illustrated using QLORA,” IEEE Access, 2024

work page 2024

[7] [7]

Multi-label classification with generative large language models,

N. Niraula, S. Ayhan, B. Chidambaram, and D. Whyatt, “Multi-label classification with generative large language models,” in Proc. of the 43rd Digital Avionics Systems Conference (DASC) , 2024

work page 2024

[8] [8]

A fine- tuned BERT-based transfer learning approach for text classification,

R. Qasim, W. H. Bangyal, M. A. Alqarni, and A. A. Almazroi, “A fine- tuned BERT-based transfer learning approach for text classification,” Journal of Healthcare Engineering , 2022

work page 2022