Pith. sign in

REVIEW 3 major objections 5 minor 47 references

CLaDMoP: Learning Transferrable Models from Successful Clinical Trials via LLMs

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Pre-training on successful clinical trials improves trial outcome prediction, especially for early phases and new diseases.

desk verdict A useful CLIP-style pre-training recipe and new dataset for clinical trial outcome prediction, with real gains on the TOP benchmark, but the transfer claim is likely inflated by SCT pre-training exposure and the abstract overstates the improvements. read the letter →

arxiv 2505.18527 v1 pith:2NF6L52V submitted 2025-05-24 cs.LG cs.AI

Yiqing Zhang , Xiaozhong Liu , Fabricio Murai This is my paper
classification cs.LGcs.AI
keywords clinicaltrialoutcomepredictionself-supervisedpre-trainingcontrastivelearningpair-matchingproxytaskBioGPTdrug–diseaserepresentationmulti-levelfeaturefusionparameter-efficientfine-tuning
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 argues that clinical-trial outcome models trained end-to-end with task-specific losses on small, phase-specific datasets overfit and fail on rare or new diseases. To fix this, the authors introduce CLaDMoP, which pre-trains on a new dataset of trials they label as successful, using a task-agnostic pair-matching objective that aligns eligibility-criteria text with drug–disease pairs. After lightweight fine-tuning, the model improves PR-AUC and ROC-AUC over the prior state of the art while matching F1, with the largest gains on phase I and phase II trials and on disease combinations never seen during training. If correct, this gives the field a transferable pre-trained predictor for trial outcomes that does not rely on task-specific objectives.

What carries the argument

The central mechanism is a two-branch contrastive architecture: a frozen BioGPT encoder processes eligibility criteria while a lightweight Drug-Molecule branch jointly encodes drug SMILES and ICD-10 disease codes, and a symmetric InfoNCE loss aligns successful pairs against all other combinations. Multi-level fusion transfers intermediate BioGPT embeddings into the Drug-Molecule branch through grouping blocks, where trainable centroid tokens aggregate information via cross-attention and progressively halve token counts, keeping computational cost manageable. In fine-tuning, only the prediction head and LoRA layers are trained, preserving the pre-trained representations.

What would settle it

Take a random sample of trials from the SCT dataset and compare their automatically assigned success labels against the actual outcome results recorded in ClinicalTrials.gov; if a meaningful fraction of labeled-successful trials are actually failed, the pre-training signal is corrupted and the claimed transfer gains would need to be re-explained.

Watch

Extended reading notes

Core claim

CLaDMoP establishes that pre-training on successful clinical trials with a task-agnostic pair-matching proxy yields transferable representations that improve clinical trial outcome prediction. Concretely, on the TOP benchmark, CLaDMoP achieves up to 10.5% higher PR-AUC and 3.6% higher ROC-AUC than MEXA-CTP while matching its F1 score, and on new-disease subsets it beats MEXA-CTP by 13.63% in phase I and 8.02% in phase II accuracy. The paper attributes this to the task-agnostic pre-training objective, which forces the model to learn generalizable joint embeddings of eligibility criteria, drug molecules, and target diseases, rather than optimizing a phase-specific classification loss.

Load-bearing premise

The entire pre-training dataset rests on the assumption that a drug reaching the market implies all its earlier-phase trials succeeded, and that reaching a later phase means all previous phases succeeded; if this heuristic mislabels many trials, the pre-trained model is trained on wrong positive signals.

Editorial extensions

If this is right

  • If the central claim holds, clinical trial outcome prediction can be improved for early-phase trials and for diseases with little or no labeled data, where task-specific supervised models typically struggle.
  • The pair-matching pre-training objective could be reused for other sparse clinical tasks, such as predicting trial enrollment, safety signals, or post-market adverse events, without redesigning the loss.
  • The grouping-block fusion design implies that intermediate LLM embeddings carry useful signal beyond the final layer, suggesting that deep features can be distilled into lightweight models at low cost.
  • Because the model matches MEXA-CTP's F1 while improving PR-AUC and ROC-AUC, it points toward a better precision-recall trade-off for prioritizing which trials to fund or run.
  • The approach strengthens the case for self-supervised pre-training as a general alternative to task-specific losses in high-stakes biomedical prediction.

Reading between the lines

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

  • A direct audit of SCT labels against recorded trial outcomes would clarify whether the gains come from the pair-matching objective itself or from the noisy success-label heuristic; if many labeled 'successful' trials actually failed, the reported transfer might shrink.
  • The same architecture could apply to other domains with one dense text modality and one structured modality, such as medical device approvals or drug repurposing, where successful cases are known but failure labels are scarce.
  • Ablating the multi-level fusion to a single LLM embedding level would isolate how much of the performance gain comes from the grouping-block design versus the contrastive objective.
  • Testing the pre-trained model in a strict zero-shot retrieval setting, without any fine-tuning, would reveal whether the representation itself transfers or whether the LoRA fine-tuning is the main driver of the reported improvements.
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 CLaDMoP, a two-stage framework for clinical trial outcome prediction: a pre-training stage that pairs frozen BioGPT embeddings of eligibility criteria with a lightweight disease-molecule branch using an InfoNCE-style contrastive loss on a newly constructed Successful Clinical Trials (SCT) dataset, followed by a parameter-efficient fine-tuning stage with LoRA and a prediction head. The authors evaluate on the TOP benchmark, comparing against classical ML baselines, HINT, and MEXA-CTP, and report improvements in PR-AUC and ROC-AUC, plus a claim of superior accuracy on test trials with disease combinations not seen in TOP training ('new diseases'). The paper also presents ablations on grouping-block configuration, LoRA variants, and data-usage strategies.

Significance. If the claims hold, the paper offers a useful pre-trained representation for trial outcome prediction, a new dataset (SCT), and an efficient grouping mechanism to fuse LLM features at multiple levels. The release of code and dataset is a concrete contribution to reproducibility. However, the headline transfer result on 'new diseases' rests on a pre-training-dataset construction choice that may introduce a confound, and the reported performance gains are not substantiated by significance testing. The core TOP benchmark results are potentially sound, but the central claim of task-agnostic transfer to new diseases needs additional controls before it can be accepted.

major comments (3)
  1. [Section 3.2 and Table 3] The new-disease transfer claim is confounded by pre-training exposure. Section 3.2 states that only TOP test trial identifiers (NCTids) are excluded from SCT, not disease combinations. Since SCT contains 3,326 diseases and 951 unique disease combinations (Table 1), it is likely that many of the 138, 340, and 198 new-disease test trials (Table 8) involve disease combinations that appeared in SCT pre-training. MEXA-CTP has no SCT pre-training, so the comparison in Table 3 is not controlled: the reported 13.63% (phase I) and 8.02% (phase II) accuracy gains could reflect data overlap rather than task-agnostic transfer. I request an analysis that excludes from SCT all trials whose disease combinations overlap with any TOP test disease combination, or an evaluation restricted to disease combinations never seen in SCT.
  2. [Section 3.2, Dataset construction] The success-labeling heuristic is a load-bearing premise for the entire pre-training stage. The assertion that 'for drugs already on the market, we assume that all clinical trials from phase I to phase III were successful' and that progression to phase p implies successful completion of all previous phases is not correct in general: many approved drugs have failed earlier- or later-phase trials, and phase I trials often assess safety rather than efficacy. If a nontrivial fraction of SCT labels are wrong, the contrastive pre-training objective is trained on noisy success signals, and the downstream gains may not generalize. The paper should validate the heuristic by comparing SCT labels with actual outcome statuses from ClinicalTrials.gov (where available) or from a source such as TrialProbe or aggregate clinical trial results, and report the resulting label-noise rate.
  3. [Section 4.2 and Table 2] The statistical support for the headline improvements is insufficient. Table 2 reports bootstrap means and standard deviations, but no significance tests or confidence intervals are given; several phase II/III ROC-AUC and F1 differences overlap at one standard deviation (e.g., phase II ROC-AUC 0.647±0.018 vs 0.638±0.005; phase III F1 0.861±0.014 vs 0.857±0.007). Moreover, the abstract and Section 6 report 'up to 10.5% and 3.6% improvement' without clarifying that these are relative percentages of the unweighted average over phases, not absolute percentage points, and the term 'significantly improves' is used without a statistical test. Please report absolute differences, bootstrap CIs or paired significance tests, and clarify the metric interpretation.
minor comments (5)
  1. [Table 3] The header 'MECA-CTP' is a typo and should read 'MEXA-CTP'.
  2. [Section 5, Related Work] The paragraph 'Large Language Models for Clinical Trials' contains a duplicated sentence: 'There is significant interest in leveraging Large Language Models (LLMs) for various prediction tasks in the context for clinical trials [46].' appears twice verbatim.
  3. [Algorithm 1] Line 6 of Algorithm 1 writes 'logits = exp(τ)· f_C × f_DM^T', which is inconsistent with standard InfoNCE temperature scaling; the exponential should be applied to the scaled similarity, not to the temperature itself. Please clarify the exact computation.
  4. [Table 6] The entry for 'w/o pre-training' in the PR-AUC column is '0.766 ± .12', which lacks a leading zero before the decimal and differs in formatting from the other entries; please standardize.
  5. [Section 4.4] The complexity analysis would be clearer if it explicitly stated the dependence of the sequence length on the number of grouping blocks and defined n after the grouping reduction; as written, the claim that the DM branch 'scales quadratically' is contrasted with the fixed centroid count of 25 without a formal derivation.

Circularity Check

1 steps flagged · score 5.0 of 10

New-disease transfer claim is partially circular: 'New Diseases' are defined only against TOP training sets, while SCT pre-training may contain the same disease combinations, so the reported 13.63%/8.02% gains are not a controlled prediction of unseen diseases.

  1. fitted input called prediction [Section 3.2 (Dataset) and Section 4.2 (New Diseases); Tables 1 and 3]
    "To prevent data leakage, we ensure that none of the trials in TOP’s test set are included in SCT’s pre-training data. Specifically, we perform a temporal split based on NCTid for train-test separation, analogous to the split done by previous works for TOP. ... We select disease combinations D(j) that appear ONLY in test sets but not in ANY training set, referring to them as 'New Diseases'."

    The 'New Diseases' subset is defined only against TOP training sets, while CLaDMoP is pre-trained on SCT before fine-tuning. The leakage prevention excludes TOP test trials by NCTid, not by disease combination. Since SCT contains 3,326 diseases and 951 unique disease combinations (Table 1), the 138/340/198 'new disease' test trials can have disease codes already present in SCT. MEXA-CTP is not pre-trained on SCT, so the comparison is uncontrolled: the phase I 13.63% and phase II 8.02% accuracy gains can reflect SCT exposure to those disease codes rather than task-agnostic transfer. In other words, the 'prediction' on new diseases is fitted to pre-training inputs that the paper's exclusion rule does not remove, making the transfer claim circular with respect to the definition of 'new'.

full rationale

CLaDMoP's central pipeline (pre-train with InfoNCE pair matching on SCT, fine-tune with class-weighted BCE on TOP) is not circular in the strict sense: the pre-training objective is a pair-matching proxy, not the downstream classification loss, and the main TOP benchmark results in Table 2 are reported on test trials excluded from SCT by NCTid. The headline 10.5%/3.6% PR-AUC/ROC-AUC improvements therefore have independent content. However, the 'New Diseases' transfer result in Table 3 is partially circular. The subset is defined as disease combinations absent from 'ANY training set' (Section 4.2), but the leakage prevention in Section 3.2 only removes TOP test trials by NCTid from SCT; it does not remove their disease combinations. Because SCT contains thousands of diseases and hundreds of disease combinations, many of the 'new disease' test trials may have disease codes present in SCT, so MEXA-CTP (with no SCT pre-training) is not a controlled comparator. The phase I 13.63% and phase II 8.02% accuracy gains on 'new diseases' are therefore not demonstrated to be task-agnostic generalization; they are at least partly explained by the model's exposure to those disease codes during pre-training. The paper's self-citations to MEXA-CTP are not load-bearing in a circular way: MEXA-CTP is used as a baseline and its encoding is reused for ML baselines, but the CLaDMoP results are not derived from MEXA-CTP's outputs or from a uniqueness theorem. No other circular steps were identified.

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

The paper does not postulate new physical or biological entities. It introduces model components (grouping blocks, trainable centroids) that are architectural, not invented entities in the ledger sense. The main uncharged assumptions are the SCT success-labeling heuristics and the proxy-task validity.

free parameters (5)
  • InfoNCE temperature tau = 0.6
    Chosen via hyperparameter tuning; scales all logits in the pre-training loss (Section 4.1).
  • Pre-training batch size = 128
    Selected from a batch-size sweep (Appendix C.3); contrastive learning is sensitive to batch size.
  • Learning rates for PEFT = 1e-2 (head), 5e-2 (LoRA phases I/II), 1e-3 (LoRA phase III)
    Set per phase; not justified beyond tuning.
  • Number of centroids per grouping block = 100->50->25
    Progressive halving; the final count of 25 is a design choice that controls token length.
  • Grouping blocks and layers (G=3, S=2) = G=3, S=2
    Ablated in Table 4; best on phase III validation, but other phases were not ablated.
assumptions (6)
  • domain assumption Drugs already on the market had successful phase I-III trials
    SCT dataset construction, Section 3.2: this labels all phases of marketed drugs as successful without per-trial evidence.
  • domain assumption A drug progressing to phase p in {II, III, IV} has successfully completed all previous phases
    SCT dataset construction, Section 3.2: used to infer success labels for non-marketed drugs.
  • ad hoc to paper Cosine similarity in the joint embedding space is a valid proxy for clinical trial success
    The pair-matching proxy task (Section 3.2) optimizes InfoNCE over this similarity; there is no independent evidence that alignment of criteria and drug-disease embeddings corresponds to true success.
  • domain assumption BioGPT's intermediate layers provide useful multi-level representations of eligibility criteria
    Multi-level fusion in Section 3.1.1 assumes that coarse, medium, and fine embeddings add signal beyond the final layer.
  • domain assumption SMILES segments encoded by DeepChem capture molecular identity
    Drug molecule embedding, Appendix A.1, assumes segment-level features from DeepChem are sufficient.
  • domain assumption ICD-10 hierarchy via icdcodex captures disease relatedness
    Target disease embedding, Appendix A.2, uses ICD-10 codes as the disease representation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLaDMoP: Learning Transferrable Models from Successful Clinical Trials via LLMs." pith.science (2026). https://pith.science/paper/2NF6L52V

@misc{pith2026250518527,
  author       = {Pith},
  title        = {Pith review of: CLaDMoP: Learning Transferrable Models from Successful Clinical Trials via LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2NF6L52V}},
  note         = {Machine review of arXiv:2505.18527}
}
read the original abstract

Many existing models for clinical trial outcome prediction are optimized using task-specific loss functions on trial phase-specific data. While this scheme may boost prediction for common diseases and drugs, it can hinder learning of generalizable representations, leading to more false positives/negatives. To address this limitation, we introduce CLaDMoP, a new pre-training approach for clinical trial outcome prediction, alongside the Successful Clinical Trials dataset(SCT), specifically designed for this task. CLaDMoP leverages a Large Language Model-to encode trials' eligibility criteria-linked to a lightweight Drug-Molecule branch through a novel multi-level fusion technique. To efficiently fuse long embeddings across levels, we incorporate a grouping block, drastically reducing computational overhead. CLaDMoP avoids reliance on task-specific objectives by pre-training on a "pair matching" proxy task. Compared to established zero-shot and few-shot baselines, our method significantly improves both PR-AUC and ROC-AUC, especially for phase I and phase II trials. We further evaluate and perform ablation on CLaDMoP after Parameter-Efficient Fine-Tuning, comparing it to state-of-the-art supervised baselines, including MEXA-CTP, on the Trial Outcome Prediction(TOP) benchmark. CLaDMoP achieves up to 10.5% improvement in PR-AUC and 3.6% in ROC-AUC, while attaining comparable F1 score to MEXA-CTP, highlighting its potential for clinical trial outcome prediction. Code and SCT dataset can be downloaded from https://github.com/murai-lab/CLaDMoP.

Figures

Figures reproduced from arXiv: 2505.18527 by the authors.

Figure 1
Figure 1. (a) Recent work has focused on using a lightweight encoder model and a prediction head trained with a task￾specific loss due to limitations in the availability of labeled data. (b) In contrast, CLaDMoP follows a two-stage train￾ing paradigm. It incorporates a large language model (LLM branch) and a lightweight attention model (Disease-Molecule branch), sharing knowledge from the former to the latter. (b-left) Pre-tr… view at source ↗
Figure 2
Figure 2. (Left) CLaDMoP consists of two branches: the LLM branch and the DM branch. The prediction head is a three-layer residual network, which is trained during the PEFT stage. Further details are provided in Section 3.1. (Right-top) The grouping block enables the model to fuse information from the LLM branch. Each grouping block consists of 𝐺 grouping layer(s), where the number of centroids is progressively reduced at eac… view at source ↗
Figure 3
Figure 3. Few-shot learning results for clinical trial outcome [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Few-shot learning results for clinical trial outcome [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Pre-trained models performance using SCT [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: (PEFT) fine-tuned models performance for three [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Batch size vs. validation loss. We show the corre [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 35 canonical work pages

  1. [1]

    Siddharth Biswal, Cao Xiao, Lucas M Glass, Elizabeth Milkovits, and Jimeng Sun. 2020. Doctor2vec: Dynamic doctor representation learning for clinical trial recruitment. In AAAI

    work page 2020

  2. [2]

    Wendy M Bjornson-Benson, Thomas B Stibolt, Kenneth A Manske, Kathleen J Zavela, Diana J Youtsey, and A Sonia Buist. 1993. Monitoring recruitment effec- tiveness and cost in a clinical trial. Controlled clinical trials 14, 2 (1993), 52–67

    work page 1993

  3. [3]

    Daniel Bobo, Kye J Robinson, Jiaul Islam, Kristofer J Thurecht, and Simon R Corrie. 2016. Nanoparticle-based medicines: a review of FDA-approved materials and clinical trials to date. Pharmaceutical research 33 (2016), 2373–2387

    work page 2016

  4. [4]

    Changyou Chen, Jianyi Zhang, Yi Xu, Liqun Chen, Jiali Duan, Yiran Chen, Son Tran, Belinda Zeng, and Trishul Chilimbi. 2022. Why do we need large batch- sizes in contrastive learning? a gradient-bias perspective. Advances in Neural Information Processing Systems 35 (2022), 33860–33875

    work page 2022

  5. [5]

    Shein-Chung Chow and Annpey Pong. 1998. An overview of the regulatory approval process in drug development. Drug information journal 32, 1_suppl (1998), 1175S–1185S

    work page 1998

  6. [6]

    Joseph A DiMasi, Ronald W Hansen, and Henry G Grabowski. 2003. The price of innovation: new estimates of drug development costs.Journal of health economics 22, 2 (2003), 151–185

    work page 2003

  7. [7]

    Nicholas S Downing, Jenerius A Aminawung, Nilay D Shah, Harlan M Krumholz, and Joseph S Ross. 2014. Clinical trial evidence supporting FDA approval of novel therapeutic agents, 2005-2012. Jama 311, 4 (2014), 368–377

    work page 2014

  8. [8]

    Sarah A Dugger, Adam Platt, and David B Goldstein. 2018. Drug development in the era of precision medicine. Nature reviews Drug discovery 17, 3 (2018), 183–196

    work page 2018

Show all 47 references
  1. [9]

    Harold Edgar and David J Rothman. 2018. New rules for new drugs: the challenge of AIDS to the regulatory process. In AIDS: Society, Ethics and Law . Routledge, 447–478

    2018

  2. [10]

    Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. 2023. Improving clip training with language rewrites. Advances in Neural Information Processing Systems 36 (2023), 35544–35575

    2023

  3. [11]

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic meta- learning for fast adaptation of deep networks. In International conference on machine learning. PMLR, 1126–1135

    2017

  4. [12]

    Tianfan Fu, Kexin Huang, Cao Xiao, Lucas M Glass, and Jimeng Sun. 2022. Hint: Hierarchical interaction network for clinical-trial-outcome predictions. Patterns 3, 4 (2022)

    2022

  5. [13]

    Junyi Gao, Cao Xiao, Lucas M Glass, and Jimeng Sun. 2020. COMPOSE: Cross- modal pseudo-siamese network for patient trial matching. In KDD. 803–812

    2020

  6. [14]

    Luyu Gao, Yunyi Zhang, Jiawei Han, and Jamie Callan. 2021. Scaling deep contrastive learning batch size under memory limited setup. arXiv preprint arXiv:2101.06983 (2021)

    2021 arXiv

  7. [15]

    Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. 2024. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608 (2024)

    2024 arXiv

  8. [16]

    Zhen-Yu Hong, Jooyong Shim, Woo Chan Son, and Changha Hwang. 2020. Pre- dicting successes and failures of clinical trials with an ensemble LS-SVR.medRxiv (2020)

    2020

  9. [17]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)

    2021 arXiv

  10. [18]

    Mingzhi Hu, Xin Zhang, Yanhua Li, Yiqun Xie, Xiaowei Jia, Xun Zhou, and Jun Luo. 2024. Only Attending What Matter within Trajectories–Memory-Efficient Trajectory Attention. In Proceedings of the 2024 SIAM International Conference on Data Mining (SDM). SIAM, 481–489

    2024

  11. [19]

    Mingzhi Hu, Zhuoyun Zhong, Xin Zhang, Yanhua Li, Yiqun Xie, Xiaowei Jia, Xun Zhou, and Jun Luo. 2023. Self-supervised pre-training for robust and generic spatial-temporal representations. In 2023 IEEE International Conference on Data Mining (ICDM). IEEE, 150–159

    2023

  12. [20]

    Ruofan Hu, Dongyu Zhang, Dandan Tao, Huayi Zhang, Hao Feng, and Elke Rundensteiner. 2023. Uce-fid: Using large unlabeled, medium crowdsourced- labeled, and small expert-labeled tweets for foodborne illness detection. In 2023 IEEE International Conference on Big Data (BigData) ...

    2023

  13. [21]

    Qiao Jin, Chuanqi Tan, Mosha Chen, Xiaozhong Liu, and Songfang Huang. 2020. Predicting Clinical Trial Results by Implicit Evidence Integration. arXiv preprint arXiv:2010.05639 (2020)

    2020 arXiv

  14. [22]

    Sravya Kakumanu, Braden J Manns, Sophia Tran, Terry Saunders-Smith, Brenda R Hemmelgarn, Marcello Tonelli, Ross Tsuyuki, Noah Ivers, Danielle Southern, Jeff Bakal, et al. 2019. Cost analysis and efficacy of recruitment strategies used in a large pragmatic community-based clini...

    2019

  15. [23]

    Richard L Kravitz, Naihua Duan, and Joel Braslow. 2004. Evidence-based medicine, heterogeneity of treatment effects, and the trouble with averages. The Milbank Quarterly 82, 4 (2004), 661–687

    2004

  16. [24]

    Robert J Levine. 1988. Ethics and regulation of clinical research . Yale University Press

    1988

  17. [25]

    Yuhang Liu, Yingxue Zhang, Xin Zhang, Ling Tian, Xu Zheng, Yanhua Li, and Jun Luo. 2025. UrbanMind: Urban Dynamics Prediction with Multifaceted Spatial- Temporal Large Language Models. arXiv preprint arXiv:2505.11654 (2025)

    2025 arXiv

  18. [26]

    Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon, and Tie-Yan Liu. 2022. BioGPT: generative pre-trained transformer for biomedical text generation and mining. Briefings in bioinformatics 23, 6 (2022), bbac409

    2022

  19. [27]

    Thamsanqa Mlotshwa, Heinrich van Deventer, and Anna Sergeevna Bosman

  20. [28]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    2018 arXiv

  21. [29]

    Youran Qi and Qi Tang. 2019. Predicting phase 3 clinical trial results by modeling phase 2 clinical trial subject level data using deep learning. In MLHC

    2019

  22. [30]

    Jianing Qiu, Wu Yuan, and Kyle Lam. 2024. The application of multimodal large language models in medicine. The Lancet Regional Health–Western Pacific (2024)

    2024

  23. [31]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  24. [32]

    Aniruddh Raghu, Maithra Raghu, Samy Bengio, and Oriol Vinyals. 2019. Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157 (2019)

    2019 arXiv

  25. [33]

    Bharath Ramsundar. 2018. Molecular machine learning with DeepChem . Ph. D. Dissertation. Stanford University

    2018

  26. [34]

    Chiman Salavati, Shannon Song, Willmar Sosa Diaz, Scott A Hale, Roberto E Montenegro, Fabricio Murai, and Shiri Dori-Hacohen. 2024. Reducing biases towards minoritized populations in medical curricular content via artificial intel- ligence for fairer health outcomes. In Procee...

    2024

  27. [35]

    Neil Savage. 2023. Drug discovery companies are customizing ChatGPT: here’s how. Nature Biotechnology (2023)

    2023

  28. [36]

    Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. 2023. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)

    2023 arXiv

  29. [37]

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. 2023. Stanford alpaca: An instruction-following llama model

    2023

  30. [38]

    Craig A Umscheid, David J Margolis, and Craig E Grossman. 2011. Key concepts of clinical trials: a narrative review. Postgraduate medicine (2011)

    2011

  31. [39]

    Cliff Wong, Sheng Zhang, Yu Gu, Christine Moung, Jacob Abel, Naoto Usuyama, Roshanthi Weerasinghe, Brian Piening, Tristan Naumann, Carlo Bifulco, and Hoifung Poon. 2023. Scaling Clinical Trial Matching Using Large Language Models: A Case Study in Oncology. In MLHC

    2023

  32. [40]

    Yang Wu, Huayi Zhang, Yizheng Jiao, Lin Ma, Xiaozhong Liu, Jinhong Yu, Dongyu Zhang, Dezhi Yu, and Wei Xu. 2024. ROSE: A Reward-Oriented Data Selection Framework for LLM Task-Specific Instruction Tuning. arXiv preprint arXiv:2412.00631 (2024)

    2024 arXiv

  33. [41]

    Jiayi Yuan, Ruixiang Tang, Xiaoqian Jiang, and Xia Hu. 2023. Large language models for healthcare data augmentation: An example on patient-trial matching. In AMIA Annual Symposium Proceedings

    2023

  34. [42]

    Huayi Zhang, Lei Cao, Samuel Madden, and Elke Rundensteiner. 2021. Lancet: labeling complex data at scale. Proceedings of the VLDB Endowment 14, 11 (2021)

    2021

  35. [43]

    Xin Zhang, Yanhua Li, Ziming Zhang, Christopher Brinton, Zhenming Liu, Zhi-Li Zhang, Hui Lu, and Zhihong Tian. 2021. Stabilized Likelihood-based Imitation Learning via Denoising Continuous Normalizing Flow. (2021)

    2021

  36. [44]

    Xingyao Zhang, Cao Xiao, Lucas M Glass, and Jimeng Sun. 2020. DeepEnroll: patient-trial matching with deep embedding and entailment prediction. In Web Conference

    2020

  37. [45]

    Yiqing Zhang, Xiaozhong Liu, and Fabricio Murai. 2025. MEXA-CTP: Mode Experts Cross-Attention for Clinical Trial Outcome Prediction. arXiv preprint arXiv:2501.06823 (2025)

    2025 arXiv

  38. [46]

    YES” or “NO

    Ailin Zhao and Yijun Wu. 2023. Future implications of ChatGPT in pharmaceutical industry: drug discovery and development. Frontiers in Pharmacology (2023). CLaDMoP: Learning Transferrable Models from Successful Clinical Trials via LLMs KDD’25, August 03–07, 2025, Toronto, Cana...

    2023

  39. [2022]

    In SACAIR

    Cauchy Loss Function: Robustness Under Gaussian and Cauchy Noise. In SACAIR

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.