Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Towards Domain Specification of Embedding Models in Medicine

T0 review · 3 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 fine-tuning a general text-embedding model with self-supervised contrastive learning on 2,033,800 medical text pairs produces embeddings that outperform both medical-domain and general-purpose baselines across a new…

desk verdict Useful new medical embedding benchmark, but the model's SOTA claim is not yet supported because train/benchmark separation is asserted, not demonstrated. read the letter →

arxiv 2507.19407 v2 pith:5S5RLRU5 submitted 2025-07-25 cs.CL

Mohammad Khodadad , Ali Shiraee Kasmaee , Mahdi Astaraki , Hamidreza Mahyar This is my paper · ORCID
classification cs.CL
keywords medicaltextembeddingscontrastivelearningbenchmarkMedTEBTEdomainadaptationclinicalNLPevaluation
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

Medical text embedding models are usually trained on a narrow slice of biomedical data and evaluated on scattered, small benchmarks, so it is hard to tell whether domain-specific models actually beat general-purpose ones. This paper claims that fine-tuning a general text-embedding model (GTE) with self-supervised contrastive learning on 2,033,800 medical text pairs yields embeddings that outperform both medical-domain baselines and leading general-purpose models, reaching an average score of 0.578 on the paper's own benchmark. To make that case, the paper introduces MedTEB, a suite of 51 tasks covering classification, clustering, pair classification, and retrieval, drawn from PubMed, MIMIC-IV, ClinicalTrials.gov, Wikipedia, and medical QA sets. If the result holds, it would mean that a single-stage, contrastive-only recipe is enough to specialise an embedding model to medicine, and that a dedicated medical benchmark is needed to expose the differences between models.

What carries the argument

The central object is the InfoNCE contrastive training objective operating on structurally paired medical texts. MedTE is created by fine-tuning GTE-Base with this objective on 2,033,800 positive pairs, where each batch is drawn from a single corpus so that all in-batch negatives are from the same data source. The same pairing structure (title-to-first-abstract-sentence, chief-complaint-to-history-of-present-illness, question-to-answer, query-to-passage) defines both the training signal and the benchmark's pair-classification and retrieval tasks, making the benchmark a direct test of whether the learned geometry separates relevant from irrelevant medical text.

What would settle it

A near-duplicate analysis between the MedTEB test items and MedTE's 2,033,800 training pairs would settle the central claim: if a substantial fraction of benchmark queries or documents have near-identical matches in the training corpus, the reported margin over out-of-distribution baselines would be an artifact of data overlap, not of contrastive domain adaptation.

Watch

Extended reading notes

Core claim

The paper's central discovery is that domain-adaptive contrastive training is decisive for medical text representation. Starting from GTE-Base, the authors fine-tune every parameter with the InfoNCE objective on positive pairs harvested from ten medical corpora—title-abstract pairs, question-explanation pairs, clinical-note snippets, and similar—and evaluate the result on MedTEB, a benchmark they construct from the same families of sources. On the benchmark, MedTE attains an average score of 0.578, ahead of the nearest competitor (MedEmbed Base, 0.539) and of the strongest general-purpose model (GTE Base, 0.529), and it leads in each of the four task families: classification (F1 0.72), clustering (V-measure 0.38), pair classification (F1 0.74), and retrieval (nDCG@10 0.45). The paper further argues that contrastive learning, not domain pretraining by itself, is the factor that separates strong from weak medical embeddings, since non-contrastive clinical models such as ClinicalBERT and BioClinicalBERT trail even general-purpose encoders.

Load-bearing premise

The load-bearing premise is that the benchmark data were fully removed from the training data, so that every score on MedTEB reflects true generalisation rather than memorisation; the paper asserts this separation but provides no deduplication analysis to verify it.

Editorial extensions

If this is right

  • If the claim holds, clinicians building retrieval or QA systems on medical text should prefer a contrastively fine-tuned embedding model over off-the-shelf clinical transformers, which underperform even general encoders.
  • MedTEB itself becomes the reference point: any future medical embedding model can be compared on the same 51 tasks, removing the current fragmentation where each paper uses its own limited benchmark.
  • The per-source results imply that gains are concentrated on clinical narrative (MIMIC-IV) and structured trial text, where MedTE's margin is largest, while all models saturate on MedMCQA-style tasks.
  • The result suggests that a single-stage contrastive fine-tune, without an additional masked-language-modeling phase, is sufficient to specialise a general embedding model to medicine.

Reading between the lines

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

  • If the benchmark truly is disjoint from training, the method's margin on MedTEB may still overstate real-world gains because MedTEB's retrieval and pair-classification tasks reuse GPT-4o paraphrases as queries; a harder test would use naturally occurring clinical questions rather than generated paraphrases.
  • The dependency on GPT-4o for both positive and negative pair construction (in pair classification and retrieval) means the benchmark's difficulty ceiling is partially set by the generator's weaknesses; a model that mirrors GPT-4o's paraphrase style could get an inflated score.
  • A natural extension of the paper's logic is to test whether the contrastive-only recipe transfers to other specialised domains, such as chemistry or law, where the same benchmark-plus-finetune pattern may hold.
  • The absence of a published deduplication analysis means the community could add a near-duplicate check between MedTEB and MedTE's training data; if overlap is discovered, the evaluation protocol would need to change for all future comparisons.
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.

Referee Report

3 major / 5 minor

Summary. The paper introduces MedTE, a GTE-base embedding model fine-tuned with self-supervised contrastive learning on 2,033,800 medical text pairs assembled from PubMed, bioRxiv/medRxiv, MIMIC-IV, ClinicalTrials.gov, MedMCQA, MedQA, MedQuAD, TREC-COVID, NFCorpus, and CURE-V1. It also introduces MedTEB, a benchmark of 51 tasks in classification, clustering, pair classification, and retrieval, constructed from overlapping data sources. The authors report that MedTE achieves an average MedTEB score of 0.578, outperforming the nearest competitor (0.539 for MedEmbed Base by one aggregation, 0.525 for Nomic Embed V1 Unsupervised by another), and conclude that MedTE establishes a new state of the art for medical text embeddings. The paper also analyzes per-source performance, training dynamics, and runtime-effectiveness trade-offs.

Significance. The paper is potentially valuable as a community resource: it releases a fine-tuned medical embedding model, a 51-task benchmark, and code repositories, and it follows a simple, reproducible self-supervised contrastive training recipe. If the reported gains are genuine, the work would demonstrate that domain-specific contrastive fine-tuning of a strong general-purpose backbone can improve medical text embeddings across multiple task families. However, the central SOTA claim is not yet supported: the benchmark and training corpus are drawn from the same sources, no overlap analysis is provided, and the appendix indicates that MedTEB scores were monitored during training. These issues must be resolved before the empirical conclusions can be accepted.

major comments (3)
  1. [§3.1, §3.3, Table 4] The claim that "all benchmark data were removed from the training data" (end of §3.1) is asserted but not demonstrated, and it is load-bearing for the headline SOTA result in Table 4. The training corpus and MedTEB are drawn from the same sources (PubMed, MIMIC-IV, ClinicalTrials.gov, MedMCQA, MedQA, MedQuAD, bioRxiv/medRxiv), and the retrieval and pair-classification tasks are constructed by generating GPT-4o paraphrases of sentences taken from the same corpora (§3.3). No deduplication, near-duplicate analysis, or overlap statistics are reported. If a nontrivial fraction of MedTEB queries/documents has exact or near-duplicate counterparts among the 2,033,800 training pairs, then MedTE is evaluated in-distribution while every baseline is out-of-distribution, and the +0.039–0.049 margin in Table 4 may reflect memorization rather than generalization. Please add a rigorous overlap analysis (e.g., exact-match counts, MinHash or embedding-similarity near-duplicate rates, per-source breakdown) and, ideally, an evaluation on independent external medical embedding benchmarks (the MTEB medical subset, HEmTEB) to support the generalization claim.
  2. [Appendix .2 (Loss) and §3.2] The appendix reports that "the average MedTEB score peaks at around step 6000 and then declines by the time it reaches step 8000," which indicates that MedTEB was used during training to monitor downstream performance and thus to inform development decisions. This undermines the status of MedTEB as a cleanly held-out test set even if the data-separation assertion in §3.1 is true. The paper should disclose all uses of MedTEB during model development, explain how the final checkpoint (CL15, step 8000) was selected without reference to MedTEB scores, and report results for a checkpoint selected solely on training/validation loss, or use a separate development set for model selection.
  3. [§4 and §7 (Table 4 and Conclusion)] The reported margins are not supported by any statistical analysis, and the headline comparison is inconsistent across sections. Table 4 gives per-task means with standard deviations, but no significance tests, paired comparisons, or per-task win rates across the 51 tasks; several margins (classification 0.72 vs 0.70, clustering 0.38 vs 0.36, retrieval 0.45 vs 0.41) are within one standard deviation. In addition, §4.1 states that the nearest competitor is MedEmbed Base at 0.539, while §7 says "0.578 versus 0.525 for the nearest competitor" (Nomic Embed V1 Unsupervised in Table 4). The authors should state which baseline and aggregation (Avg vs AvgAll) is used for the SOTA claim and provide a statistical comparison (e.g., paired bootstrap or Wilcoxon tests, or per-task win rates) before claiming consistent superiority.
minor comments (5)
  1. [§7 Conclusion] The nearest-competitor value 0.525 in the conclusion conflicts with the 0.539 value reported in §4.1; please reconcile the two statements and clarify which column of Table 4 is being cited.
  2. [§4.4] The sentence "For semantically equivalent sentence detection, posts an F1 of 0.74±0.17, Thenlper GTE Base by+0.05 and MedEmbed Base by +0.06" is missing a subject and should be rewritten for grammatical completeness.
  3. [Table 5 and Figure 1] Naming is inconsistent: Table 4 and §3 use "MedTE," while Table 5 and Figures 1 and 3 use "Ours"; please unify the model name throughout.
  4. [Figure captions] Several figure captions contain typos, including "Perforamcne" in Figures 1 and 3, and the caption of Figure 4a refers to "MTEB Score" even though the figure reports MedTEB scores; these should be corrected.
  5. [§7 Conclusion] The conclusion lists Med-BERT among the baselines that MedTE outperforms, but Med-BERT does not appear in Table 3 or Table 4; please either add it to the evaluation or remove the mention.

Circularity Check

0 steps flagged · score 2.0 of 10

No construction-level circularity; the SOTA claim depends on an asserted but unverified train/benchmark separation, a correctness risk rather than a circular reduction.

full rationale

The claimed derivation (MedTE: InfoNCE contrastive fine-tuning of GTE-base on 2,033,800 medical text pairs; MedTEB: a 51-task benchmark) is not circular at the level of equations or definitions. The benchmark tasks are not defined in terms of MedTE's outputs, and the SOTA result is an empirical comparison against external baselines in Table 4, not a fitted parameter. The only self-citation, ChemTEB [27], is offered as inspiration for a domain-specific benchmark, not as load-bearing justification. The most plausible circularity concern is the shared provenance: Section 3.1 states 'For both model training and benchmarking, we used these data sources' and then asserts 'all benchmark data were removed from the training data' with no deduplication analysis. If that removal failed, MedTEB would be an in-distribution test for MedTE and out-of-distribution for baselines, which would invalidate the SOTA claim. Similarly, Appendix .2 reports that 'the average MedTEB score peaks at around step 6000,' indicating the benchmark was monitored during training. These are validity and rigor concerns, not construction-level circular reductions: the paper explicitly asserts the separation, and absent evidence of overlap we cannot exhibit Eq. X = Eq. Y or a fitted-input-renamed-as-prediction. Therefore the circularity score is low, with the main caveat being an unverified independence assumption.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claims rest on the training data construction, the benchmark construction, and the assertion of train/benchmark separation. None of these are externally verified, and the model and benchmark share the same source corpora, creating a self-referential evaluation. No new physical or conceptual entities are introduced.

free parameters (7)
  • InfoNCE temperature tau = learned, initial value not reported
    Temperature in Eq. 1 is learned during training; its initial value is a hidden choice that affects the contrastive loss and the resulting embedding geometry.
  • Training steps = 8000
    Chosen for convergence; the paper notes that the average MedTEB score peaks around step 6000, so the final checkpoint is not the peak benchmark performer.
  • Batch size = 1024
    Larger batches change the negative sampling distribution in InfoNCE; chosen for throughput.
  • Peak learning rate = not reported
    The learning rate schedule is described only qualitatively (warmup, cosine decay); the peak value is unstated and affects training.
  • Corpus sampling proportions = not reported
    Training pairs are 'drawn proportionally from all corpora' (Table 1), but exact proportions are not given; they determine the model's domain mix.
  • GPT-4o paraphrase generation settings = not reported
    Benchmark positive pairs are generated by prompting GPT-4o; the prompt, temperature, and number of candidates are unspecified, so the benchmark is not exactly reproducible.
  • Hard-negative sampling count = 64
    The number of nearest neighbors used to sample negatives in pair classification is a design choice that affects task difficulty and model rankings.
assumptions (5)
  • domain assumption Positive-pair heuristics (e.g., title to first abstract sentence) capture semantic equivalence
    Table 1 treats structurally related texts as semantically aligned; if these pairs are noisy, the learned embeddings are less meaningful.
  • domain assumption Benchmark data are fully disjoint from training data
    Section 3.1 asserts removal of benchmark data from training, but no deduplication analysis is provided and the same source corpora are used for both.
  • domain assumption GPT-4o paraphrases are semantically equivalent to the original texts
    Pair classification and retrieval labels are built on generated paraphrases; errors in paraphrase equivalence propagate to benchmark scores.
  • domain assumption The benchmark's data sources are representative of real-world medical text
    The benchmark draws from PubMed, MIMIC-IV, ClinicalTrials.gov, preprints, and QA sets; if these do not cover the deployment distribution, MedTEB scores may not generalize.
  • standard math GTE-base mean pooling is an appropriate encoder for medical text
    The model inherits GTE's architecture and pooling; this is a standard technique, not introduced by the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Domain Specification of Embedding Models in Medicine." pith.science (2026). https://pith.science/paper/5S5RLRU5

@misc{pith2026250719407,
  author       = {Pith},
  title        = {Pith review of: Towards Domain Specification of Embedding Models in Medicine},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5S5RLRU5}},
  note         = {Machine review of arXiv:2507.19407}
}
read the original abstract

Medical text embedding models are foundational to a wide array of healthcare applications, ranging from clinical decision support and biomedical information retrieval to medical question answering, yet they remain hampered by two critical shortcomings. First, most models are trained on a narrow slice of medical and biological data, beside not being up to date in terms of methodology, making them ill suited to capture the diversity of terminology and semantics encountered in practice. Second, existing evaluations are often inadequate: even widely used benchmarks fail to generalize across the full spectrum of real world medical tasks. To address these gaps, we leverage MEDTE, a GTE model extensively fine-tuned on diverse medical corpora through self-supervised contrastive learning across multiple data sources, to deliver robust medical text embeddings. Alongside this model, we propose a comprehensive benchmark suite of 51 tasks spanning classification, clustering, pair classification, and retrieval modeled on the Massive Text Embedding Benchmark (MTEB) but tailored to the nuances of medical text. Our results demonstrate that this combined approach not only establishes a robust evaluation framework but also yields embeddings that consistently outperform state of the art alternatives in different tasks.

Figures

Figures reproduced from arXiv: 2507.19407 by the authors.

Figure 1
Figure 1. Comparison of model performance on Wikipedia versus MIMIC-IV [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. t-SNE Visualization of Wikipedia Disease Embeddings by Body System. [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Comparison of Models’ performance on different sources of tasks in MedTEB [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of Average Score vs Time 0 2000 4000 6000 8000 10000 12000 Global training step 0.4 0.6 0.8 1.0 1.2 Loss Learning Curve: Training (512-Step Mean) and Evaluation Loss vs. Global Step Train (mean of 512 steps) Eval loss [PITH_FULL_IMAGE:figures/full_fig_p0…
Figure 5
Figure 5. Figure 5: Training and validation loss curves for M [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. VeriLLMed: Interactive Visual Debugging of Medical Large Language Models with Knowledge Graphs

    cs.CL 2026-04 unverdicted novelty 6.0 of 10

    VeriLLMed uses biomedical knowledge graphs to turn medical LLM reasoning into comparable paths and automatically flags three recurring error types: relation, branch, and missing errors.

Reference graph

Works this paper leans on

43 extracted references · 26 canonical work pages · cited by 1 Pith paper

  1. [1]

    A survey of word embeddings for clinical text

    Faiza Khan Khattak, Serena Jeblee, Chloé Pou-Prom, Mohamed Abdalla, Christopher Meaney, and Frank Rudzicz. A survey of word embeddings for clinical text. Journal of Biomedical Informatics, 100:100057, 2019

    work page 2019

  2. [2]

    Biobert: a pre-trained biomedical language representation model for biomedical text mining

    Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234–1240, 2020

    2020

  3. [3]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

    arXiv 2018

  4. [4]

    Clinicalbert: Modeling clinical notes and predicting hospital readmission

    Kexin Huang, Jaan Altosaar, and Rajesh Ranganath. Clinicalbert: Modeling clinical notes and predicting hospital readmission. arXiv preprint arXiv:1904.05342, 2019

    arXiv 1904

  5. [5]

    Med-bert: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction

    Laila Rasmy, Yang Xiang, Ziqian Xie, Cui Tao, and Degui Zhi. Med-bert: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction. NPJ digital medicine, 4(1):86, 2021

    work page 2021

  6. [6]

    Text embeddings by weakly-supervised contrastive pre-training

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533, 2022

    arXiv 2022

  7. [7]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    N Reimers. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019

    arXiv 1908

  8. [8]

    Mteb: Massive text embedding benchmark

    Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers. Mteb: Massive text embedding benchmark. arXiv preprint arXiv:2210.07316, 2022

    arXiv 2022

Show all 43 references
  1. [9]

    Ehr phenotyping via jointly embedding medical concepts and words into a unified vector space

    Tian Bai, Ashis Kumar Chanda, Brian L Egleston, and Slobodan Vucetic. Ehr phenotyping via jointly embedding medical concepts and words into a unified vector space. BMC medical informatics and decision making, 18(Suppl 4):123, 2018

    2018

  2. [10]

    Automated disease cohort selection using word embeddings from electronic health records

    Benjamin S Glicksberg, Riccardo Miotto, Kipp W Johnson, Khader Shameer, Li Li, Rong Chen, and Joel T Dudley. Automated disease cohort selection using word embeddings from electronic health records. In PACIFIC SYMPOSIUM on BIOCOMPUTING 2018: Proceedings of the Pacific Symposium...

    2018

  3. [11]

    Retrieval-augmented generation for knowledge- intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge- intensive nlp tasks. Advances in neural information processing system...

    2020

  4. [12]

    Better synonyms for enriching biomedical search

    Lana Yeganova, Sun Kim, Qingyu Chen, Grigory Balasanov, W John Wilbur, and Zhiyong Lu. Better synonyms for enriching biomedical search. Journal of the American Medical Informatics Association, 27(12):1894–1902, 2020

    1902

  5. [13]

    Combining structured and unstructured data for predictive models: a deep learning approach

    Dongdong Zhang, Changchang Yin, Jucheng Zeng, Xiaohui Yuan, and Ping Zhang. Combining structured and unstructured data for predictive models: a deep learning approach. BMC medical informatics and decision making, 20(1):280, 2020

    2020

  6. [14]

    Scibert: A pretrained language model for scientific text

    Iz Beltagy, Kyle Lo, and Arman Cohan. Scibert: A pretrained language model for scientific text. arXiv preprint arXiv:1903.10676, 2019

    1903 arXiv

  7. [15]

    M3–Embedding: Multi–linguality, multi–functionality, multi–granularity text embeddings through self–knowledge distillation

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. M3–Embedding: Multi–linguality, multi–functionality, multi–granularity text embeddings through self–knowledge distillation. arXiv preprint, arXiv:2402.03216, 2024

    2024 arXiv

  8. [16]

    Towards general text embeddings with multi-stage contrastive learning

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281, 2023

    2023 arXiv

  9. [17]

    Nomic embed: Training a reproducible long context text embedder

    Zach Nussbaum, John X Morris, Brandon Duderstadt, and Andriy Mulyar. Nomic embed: Training a reproducible long context text embedder. arXiv preprint arXiv:2402.01613, 2024

    2024 arXiv

  10. [18]

    exbert: Extending pre-trained models with domain-specific vocabulary under constrained training resources

    Wen Tai, HT Kung, Xin Luna Dong, Marcus Comiter, and Chang-Fu Kuo. exbert: Extending pre-trained models with domain-specific vocabulary under constrained training resources. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1433–1439, 2020

    2020

  11. [19]

    Smith, et al

    Xi Yang, Aokun Chen, Nima PourNejatian, Hoo Chang Shin, Kaleb E. Smith, et al. Gatortron: A large clinical language model to unlock patient information from unstructured electronic health records. arXiv preprint, arXiv:2203.03540, 2022

    2022 arXiv

  12. [20]

    Biosimcse: Biomedical sentence embeddings using contrastive learning

    Kamal Raj Kanakarajan, Bhuvana Kundumani, Abhijith Abraham, and Malaikannan Sankarasubbu. Biosimcse: Biomedical sentence embeddings using contrastive learning. In Proceedings of the 13th International Workshop on Health Text Mining and Information Analysis (LOUHI), pages 81–86...

    2022

  13. [21]

    Self-supervised segment contrastive learning for medical document representation

    Waheed Ahmed Abro, Hanane Kteich, and Zied Bouraoui. Self-supervised segment contrastive learning for medical document representation. In International Conference on Artificial Intelligence in Medicine , pages 312–321. Springer, 2024

    2024

  14. [22]

    Adaption bert for medical information processing with chatgpt and contrastive learning

    Lingtong Min, Ziman Fan, Feiyang Dou, Jiaao Sun, Changsheng Luo, and Qinyi Lv. Adaption bert for medical information processing with chatgpt and contrastive learning. Electronics, 13(13):2431, 2024

    2024

  15. [23]

    Notecontrast: Contrastive language- diagnostic pretraining for medical text

    Prajwal Kailas, Max Homilius, Rahul C Deo, and Calum A MacRae. Notecontrast: Contrastive language- diagnostic pretraining for medical text. In Machine Learning for Health (ML4H) , pages 201–216. PMLR, 2023

    2023

  16. [24]

    Metadata-induced contrastive learning for zero-shot multi-label text classification

    Yu Zhang, Zhihong Shen, Chieh-Han Wu, Boya Xie, Junheng Hao, Ye-Yi Wang, Kuansan Wang, and Jiawei Han. Metadata-induced contrastive learning for zero-shot multi-label text classification. In Proceedings of the ACM Web Conference 2022, pages 3162–3173, 2022

    2022

  17. [25]

    Medembed: Medical-focused embedding models, 2024

    Abhinand Balachandran. Medembed: Medical-focused embedding models, 2024

    2024

  18. [26]

    BioLORD–2023: Semantic textual representations fusing large language models and clinical knowledge graph insights

    François Remy, Kris Demuynck, and Thomas Demeester. BioLORD–2023: Semantic textual representations fusing large language models and clinical knowledge graph insights. Journal of the American Medical Informatics Association, 31(9):1844–1855, 2024

    2023

  19. [27]

    Chemteb: Chemical text embedding benchmark, an overview of embedding models performance & efficiency on a specific domain

    Ali Shiraee Kasmaee, Mohammad Khodadad, Mohammad Arshi Saloot, Nick Sherck, Stephen Dokas, Hamidreza Mahyar, and Soheila Samiee. Chemteb: Chemical text embedding benchmark, an overview of embedding models performance & efficiency on a specific domain. arXiv preprint arXiv:2412...

    2024 arXiv

  20. [28]

    A scalable framework for benchmarking embedding models for semantic medical tasks

    Shelly Soffer, Benjamin S Glicksberg, Patricia Kovatch, Orly Efros, Robert Freeman, Alexander W Charney, Girish N Nadkarni, and Eyal Klang. A scalable framework for benchmarking embedding models for semantic medical tasks. medRxiv, pages 2024–08, 2024

    2024

  21. [29]

    Introducing hemteb: An open-source benchmark for health information retrieval, 2024

    Clinia. Introducing hemteb: An open-source benchmark for health information retrieval, 2024

    2024

  22. [30]

    Benchmarking transformer embedding models for biomedical terminology standardization

    Aditya Lahiri, Sangeeta Shukla, Ben Stear, and Taha Mohseni Ahooyi. Benchmarking transformer embedding models for biomedical terminology standardization. Machine Learning with Applications, 21:100413, 2025

    2025

  23. [31]

    Pubmed, 1996

    National Library of Medicine (US). Pubmed, 1996. Updated 30 May 2025; accessed 28 Jun 2025

    1996

  24. [32]

    Pubmed central, 2000

    National Library of Medicine (US). Pubmed central, 2000. Updated 12 Jun 2025; accessed 28 Jun 2025

    2000

  25. [33]

    Mimic-iv, a freely accessible electronic health record dataset

    Alistair EW Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J Pollard, Sicheng Hao, Benjamin Moody, Brian Gow, et al. Mimic-iv, a freely accessible electronic health record dataset. Scientific data, 10(1):1, 2023

    2023

  26. [34]

    Clinicaltrials.gov, 2000

    National Library of Medicine (US). Clinicaltrials.gov, 2000. Updated 18 Jun 2025; accessed 28 Jun 2025

    2000

  27. [35]

    biorxiv: the preprint server for biology

    Richard Sever, Ted Roeder, Samantha Hindle, and et al. biorxiv: the preprint server for biology. bioRxiv, 2019. preprint

    2019

  28. [36]

    medrxiv, 2019

    medRxiv. medrxiv, 2019. accessed 28 Jun 2025

    2019

  29. [37]

    Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering

    Ankit Pal, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering. In Conference on health, inference, and learning, pages 248–260. PMLR, 2022

    2022

  30. [38]

    What disease does this patient have? a large-scale open domain question answering dataset from medical exams

    Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Applied Sciences, 11(14):6421, 2021

    2021

  31. [39]

    A question-entailment approach to question answering

    Asma Ben Abacha and Dina Demner-Fushman. A question-entailment approach to question answering. BMC Bioinform., 20(1):511:1–511:23, 2019

    2019

  32. [40]

    Trec-covid: constructing a pandemic information retrieval test collection

    Ellen V oorhees, Tasmeer Alam, Steven Bedrick, Dina Demner-Fushman, William R Hersh, Kyle Lo, Kirk Roberts, Ian Soboroff, and Lucy Lu Wang. Trec-covid: constructing a pandemic information retrieval test collection. In ACM SIGIR Forum, volume 54, pages 1–12. ACM New York, NY , ...

    2021

  33. [41]

    A full-text learning to rank dataset for medical information retrieval

    Vera Boteva, Demian Gholipour, Artem Sokolov, and Stefan Riezler. A full-text learning to rank dataset for medical information retrieval. 2016

    2016

  34. [42]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  35. [43]

    Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters

    Jeff Rasley, Samyam Rajbhandari, Olatunji Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 3...

    2020

Pith tools

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