arxiv: 2605.06335 · v1 · submitted 2026-05-07 · 💻 cs.LG

Recognition: unknown

Eliciting associations between clinical variables from LLMs via comparison questions across populations

Fabian Kabus , Kian Kordtomeikel , Thomas Brox , Heinz Wiendl , Daiana Stolz , Harald Binder

Authors on Pith no claims yet

Pith reviewed 2026-05-08 12:56 UTC · model grok-4.3

classification 💻 cs.LG

keywords LLMsclinical variablestriplet comparisonscorrelation elicitationinvariant causal predictionCOPDmultiple sclerosis

0 comments

The pith

LLMs can recover clinical variable associations through structured patient comparison questions rather than direct queries.

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

The paper tests whether large language models contain usable information about links between patient traits by asking them indirect comparison questions instead of straightforward ones. The approach presents the model with triplets of patients and measures how similarity judgments on one trait shift when a second trait is revealed for one patient. These shifts are modeled statistically to estimate correlations, then prompt wording is varied to simulate different patient populations so that invariant causal prediction can identify a small set of stable parent links. Demonstrations in chronic obstructive pulmonary disease and multiple sclerosis produce smooth, clinically sensible correlation patterns that differ enough across prompts to support the invariance step.

Core claim

Indirect elicitation via triplet comparisons can recover meaningful association structure from LLMs and offer a cautious route from implicit correlations to causal statements that are congruent with LLM answering patterns.

What carries the argument

Patient comparison triplet questions fed to an LLM, combined with a statistical model that turns similarity decisions into correlation estimates and prompt-level environment shifts that enable invariant causal prediction across subpopulations.

If this is right

Elicited correlations are smooth, stable, and clinically interpretable across prompted environments.
Statistically significant variation across environments supports downstream invariance testing.
Invariant causal prediction yields a small set of candidate invariant parent links.
The method avoids direct elicitation while remaining congruent with the LLM's own answering patterns.

Where Pith is reading between the lines

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

The same triplet-comparison logic could be tried in non-medical domains where LLMs hold implicit structured knowledge but direct questions risk bias.
If the correlations prove robust, they might serve as a low-cost way to surface hypotheses for later verification in real patient data.
Extending the prompt environments beyond simple subpopulation shifts could test whether the invariance step still isolates plausible causal candidates.

Load-bearing premise

LLM similarity decisions on comparison triplets faithfully reflect correlations present in its training data and prompt-level environment shifts correspond to distinct subpopulations in a way that enables valid invariance testing.

What would settle it

Whether the correlations obtained from triplet comparisons on COPD and MS variables match the direction and strength of associations documented in independent clinical literature, or whether the candidate invariant parent links remain stable under alternative prompt phrasings.

Figures

Figures reproduced from arXiv: 2605.06335 by Daiana Stolz, Fabian Kabus, Harald Binder, Heinz Wiendl, Kian Kordtomeikel, Thomas Brox.

**Figure 1.** Figure 1: Main extracted associations. (a) Answer-grid response fractions with surrogate decision view at source ↗

**Figure 2.** Figure 2: Robustness checks. (a) Sensitivity to the reference value view at source ↗

**Figure 3.** Figure 3: Full COPD answer-grid matrix. 16 view at source ↗

**Figure 4.** Figure 4: Full MS answer-grid matrix. 17 view at source ↗

read the original abstract

The training data of large language models (LLMs) comprises a wide range of biomedical literature, reflecting data from many different patient populations. We investigate how it might be possible to recover information on correlation and causal links between patient characteristics, as a key building block for medical decision making. To avoid the pitfalls of direct elicitation, we propose an approach based on structured comparison questions, specifically patient comparison triplet questions. This is combined with a statistical model for the LLM representation that provides estimates of correlations without access to activations or model internals. Intuitively, we consider how similarity decisions of LLMs based on a first variable are affected by providing information on a second variable for one of the patients being assessed. We then induce prompt-level environment shifts to obtain correlation estimates for different subpopulations, which enables an invariant causal prediction (ICP) approach to obtain conservative candidate parent links. We demonstrate the method in two clinical domains, chronic obstructive pulmonary disease (COPD) and multiple sclerosis (MS). Across prompted environments, the elicited correlations are smooth, stable, and clinically interpretable, yet vary in a statistically significant way that supports downstream invariance testing, such that ICP provides a small set of candidate invariant parent links. These results show that indirect elicitation via triplet comparisons can recover meaningful association structure from LLMs and offer a cautious route from implicit correlations to causal statements that are congruent with LLM answering patterns.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Triplet comparisons let them pull stable clinical correlations from LLMs and feed them into ICP across prompt shifts, but those shifts may not give the independent environments ICP needs.

read the letter

The main thing to know is that the paper avoids direct questions to LLMs by using how the model decides similarity on patient triplets, then changes the prompt wording to stand in for different subpopulations and runs invariant causal prediction to find a few candidate parent links in COPD and MS cases. The correlations that come out are described as smooth, stable, and matching clinical expectations, with enough change across the prompted setups to support the ICP step and produce a small set of invariant candidates. That combination of triplet probing plus prompt environments for ICP is not a standard move in the LLM interpretability or causal discovery literature they cite, so the pipeline itself is the fresh element. They also use a statistical model to estimate the correlations from the similarity decisions without needing model internals, which is a practical workaround. The results stay cautious and tie back to patterns in how the LLM answers rather than claiming real-world causality. The soft spot is exactly the one in the stress-test note. Prompt modifications on a single fixed model do not automatically create the kind of independent distributional shifts that ICP requires for valid invariance testing. The abstract says the variation is statistically significant, but without seeing effect sizes, baseline comparisons, or any check that the prompted environments track actual subpopulation differences rather than shared model knowledge or wording artifacts, the candidate parents could be empty or spurious. The paper does not overclaim, which is good, but the causal step still rests on an assumption that needs more support. The work is aimed at people who want to mine LLMs for structured medical associations or test causal ideas on implicit text representations. A reader focused on interpretability methods or hypothesis generation in biomedicine would find the concrete domain results useful even if they treat the ICP output as exploratory. It deserves a serious referee because the core idea is original and applied to real clinical variables, though any review will likely press on the environment validity and ask for quantitative details that the abstract leaves out. I would send it to review.

Referee Report

2 major / 1 minor

Summary. The paper proposes an indirect elicitation method using patient comparison triplet questions to recover correlations between clinical variables from LLMs, paired with a statistical model that estimates these associations from similarity decisions without access to model internals or activations. Prompt-level environment shifts are induced to represent distinct subpopulations, enabling application of invariant causal prediction (ICP) to obtain conservative candidate causal parent links. Demonstrated in COPD and MS domains, the elicited correlations are claimed to be smooth, stable, and clinically interpretable, with statistically significant variation across environments that supports ICP yielding a small set of invariant parent candidates.

Significance. If the central claims hold, the work offers a cautious, indirect route to extract structured association and causal information from LLMs' implicit biomedical knowledge without direct prompting biases or internal access. This could provide a building block for medical decision support by leveraging the diverse patient populations reflected in LLM training data, while the ICP step adds conservatism to causal interpretations congruent with model answering patterns.

major comments (2)

[Methods (environment construction and ICP)] The application of ICP to prompt-induced environments (described in the methods for environment shifts and ICP) is load-bearing for the causal claims, yet the environments are generated solely via linguistic modifications to queries on a single fixed LLM. No demonstration is provided that these shifts produce independent changes in the implicit joint distributions of clinical variables (as required for ICP to isolate invariant mechanisms), rather than shared model artifacts or prompt-specific effects; the abstract's report of statistically significant variation does not address this.
[Abstract and Results] The abstract and results sections claim 'smooth, stable, and clinically interpretable' correlations plus 'statistically significant variation' supporting ICP, but provide no quantitative metrics, error bars, baseline comparisons, p-values, effect sizes, or details on how the parameters of the statistical model for LLM representation are estimated. This absence undermines evaluation of the evidence strength for the pipeline's outputs and the downstream invariance testing.

minor comments (1)

[Methods] The description of the triplet comparison questions and how similarity decisions translate to correlation estimates could be clarified with an explicit equation or pseudocode for the statistical model, to improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. We address each major comment point by point below, providing clarifications and indicating planned revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Methods (environment construction and ICP)] The application of ICP to prompt-induced environments (described in the methods for environment shifts and ICP) is load-bearing for the causal claims, yet the environments are generated solely via linguistic modifications to queries on a single fixed LLM. No demonstration is provided that these shifts produce independent changes in the implicit joint distributions of clinical variables (as required for ICP to isolate invariant mechanisms), rather than shared model artifacts or prompt-specific effects; the abstract's report of statistically significant variation does not address this.

Authors: We thank the referee for this important observation on the ICP application. The environments are constructed via targeted linguistic modifications to the triplet comparison prompts on a fixed LLM, as detailed in the methods, to probe subpopulation-specific associations reflected in the model's training data. The reported statistically significant variation across environments demonstrates that the shifts meaningfully affect the elicited correlations, providing a basis for invariance testing. We do not assert that these prompt shifts generate fully independent joint distributions equivalent to real-world data sources; instead, ICP is applied conservatively to identify a small set of candidate parent links that remain invariant to the induced variations, ensuring congruence with the LLM's answering behavior. In revision, we will expand the methods and discussion sections to explicitly address the assumptions of prompt-based environments, discuss potential artifacts, and add supplementary robustness checks (e.g., sensitivity to prompt phrasing and cross-LLM consistency where feasible). revision: partial
Referee: [Abstract and Results] The abstract and results sections claim 'smooth, stable, and clinically interpretable' correlations plus 'statistically significant variation' supporting ICP, but provide no quantitative metrics, error bars, baseline comparisons, p-values, effect sizes, or details on how the parameters of the statistical model for LLM representation are estimated. This absence undermines evaluation of the evidence strength for the pipeline's outputs and the downstream invariance testing.

Authors: We agree that greater quantitative detail would improve the evaluation of our claims. The full results section includes estimates from the statistical model (based on a likelihood over triplet similarity decisions), but we will revise the abstract to report specific metrics such as mean correlation strengths with standard errors, p-values for cross-environment variation, and effect sizes. The results will be augmented with error bars on all figures and tables, explicit baseline comparisons (including direct elicitation where applicable), and a clearer description of parameter estimation (including the model's formulation, optimization procedure, and any regularization). These changes will directly support the claims of smoothness, stability, interpretability, and the validity of the ICP step. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method applies external ICP to elicited LLM data

full rationale

The derivation proceeds by eliciting pairwise associations via triplet comparison questions posed to an LLM, fitting a statistical model to those responses to obtain correlation estimates, inducing prompt-based environment shifts, and then applying the external ICP procedure to identify invariant parent candidates. No equation or step reduces a claimed prediction to a fitted parameter by construction, no term is defined in terms of its output, and no load-bearing premise rests on a self-citation chain. The central result is an empirical procedure whose outputs are falsifiable against clinical knowledge and whose inputs (LLM answers) are generated independently of the ICP step. This is the normal non-circular case for a methodological proposal.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

Abstract-only review limits visibility into exact parameters and assumptions; the statistical model for LLM representation and the definition of prompt environments are likely to contain fitted components and domain assumptions about how LLMs encode population data.

free parameters (1)

parameters of the statistical model for LLM representation
The model that turns triplet answers into correlation estimates must contain parameters estimated from the LLM responses.

axioms (2)

domain assumption LLM similarity decisions on comparison questions reflect correlations in training data
Invoked when interpreting changes in answers as evidence of associations.
domain assumption prompt-level environment shifts induce distinct subpopulation environments
Required for the ICP step to be meaningful.

pith-pipeline@v0.9.0 · 5564 in / 1333 out tokens · 41489 ms · 2026-05-08T12:56:32.444425+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

56 extracted references · 27 canonical work pages · 1 internal anchor

[1]

2023 , doi =

Ban, Taiyu and Chen, Lyuzhou and Lyu, Derui and Wang, Xiangyu and Chen, Huanhuan , title =. 2023 , doi =

2023
[2]

IEEE Transactions on Artificial Intelligence , year =

Ban, Taiyu and Chen, Lyuzhou and Lyu, Derui and Wang, Xiangyu and Zhu, Qinrui and Tu, Qiang and Chen, Huanhuan , title =. IEEE Transactions on Artificial Intelligence , year =
[3]

IEEE Transactions on Knowledge and Data Engineering , year =

Ban, Taiyu and Chen, Lyuzhou and Lyu, Derui and Wang, Xiangyu and Zhu, Qinrui and Chen, Huanhuan , title =. IEEE Transactions on Knowledge and Data Engineering , year =. doi:10.1109/TKDE.2025.3528461 , issn =

work page doi:10.1109/tkde.2025.3528461 2025
[4]

Computational Linguistics , year =

Belinkov, Yonatan , title =. Computational Linguistics , year =
[5]

Felix Brandt

Bradley, Ralph Allan and Terry, Milton E. , title =. Biometrika , publisher =. 1952 , volume =. doi:10.2307/2334029 , eprint =

work page doi:10.2307/2334029 1952
[6]

org/posts/baJyjpktzmcmRfosq/ stitching-saes-of-different-sizes

Conmy, Arthur and. Towards automated circuit discovery for mechanistic interpretability , publisher =. 2023 , month =. doi:10.48550/arXiv.2304.14997 , eprint =

work page doi:10.48550/arxiv.2304.14997 2023
[7]

2024 , doi =

Darvariu, Victor-Alexandru and Hailes, Stephen and Musolesi, Mirco , title =. 2024 , doi =

2024
[8]

2021 , journal =

Elazar, Yanai and Ravfogel, Shauli and Jacovi, Alon and Goldberg, Yoav , title =. 2021 , journal =

2021
[9]

Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

Feng, Tao and Qu, Lizhen and Tandon, Niket and Li, Zhuang and Kang, Xiaoxi and Haffari, Gholamreza , title =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =
[10]

Convergent Evolution: How Different Language Models Learn Similar Number Representations

Fu, Deqing and Zhou, Tianyi and Belkin, Mikhail and Sharan, Vatsal and Jia, Robin , title =. 2026 , month =. doi:10.48550/arXiv.2604.20817 , eprint =

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.20817 2026
[11]

The Twelfth International Conference on Learning Representations , year =

Gurnee, Wes and Tegmark, Max , title =. The Twelfth International Conference on Learning Representations , year =
[12]

ACM Comput

Ji, Ziwei and Lee, Nayeon and Frieske, Rita and Yu, Tiezheng and Su, Dan and Xu, Yan and Ishii, Etsuko and Bang, Yejin and Chen, Delong and Dai, Wenliang and Chan, Ho Shu and Madotto, Andrea and Fung, Pascale , title =. ACM Computing Surveys , year =. doi:10.1145/3571730 , issn =

work page doi:10.1145/3571730
[13]

Can large language models infer causation from correlation? , booktitle =

Jin, Zhijing and Liu, Jiarui and Lyu, Zhiheng and Poff, Spencer and Sachan, Mrinmaya and Mihalcea, Rada and Diab, Mona and Sch. Can large language models infer causation from correlation? , booktitle =. 2024 , url =

2024
[14]

Advances in Neural Information Processing Systems , year =

Jin, Zhijing and Chen, Yuen and Leeb, Felix and Gresele, Luigi and Kamal, Ojasv and Lyu, Zhiheng and Blin, Kevin and Adauto, Fernando Gonzalez and. Advances in Neural Information Processing Systems , year =
[15]

Assessing multimodal chronic wound embeddings with expert triplet agreement , publisher =

Kabus, Fabian and Hindel, Julia and Bratuli. Assessing multimodal chronic wound embeddings with expert triplet agreement , publisher =. 2026 , month =. doi:10.48550/arXiv.2603.29376 , eprint =

work page doi:10.48550/arxiv.2603.29376 2026
[16]

, title =

Khatibi, Elahe and Abbasian, Mahyar and Yang, Zhongqi and Azimi, Iman and Rahmani, Amir M. , title =. 2025 , month =. doi:10.48550/arXiv.2405.01744 , eprint =

work page doi:10.48550/arxiv.2405.01744 2025
[17]

Lindsey, Jack and Gurnee, Wes and Ameisen, Emmanuel and Chen, Brian and Pearce, Adam and Turner, Nicholas L. and Citro, Craig and Abrahams, David and Carter, Shan and Hosmer, Basil and Marcus, Jonathan and Sklar, Michael and Templeton, Adly and Bricken, Trenton and McDougall, Callum and Cunningham, Hoagy and Henighan, Thomas and Jermyn, Adam and Jones, An...
[18]

Can large language models build causal graphs? , booktitle =

Long, Stephanie and Schuster, Tibor and Pich. Can large language models build causal graphs? , booktitle =. 2022 , url =

2022
[19]

Causal discovery with language models as imperfect experts , booktitle =

Long, Stephanie and Pich. Causal discovery with language models as imperfect experts , booktitle =. 2023 , url =

2023
[20]

The First Conference on Language Modeling , year =

Mallen, Alex and Brumley, Madeline and Kharchenko, Julia and Belrose, Nora , title =. The First Conference on Language Modeling , year =
[21]

The First Conference on Language Modeling , year =

Marks, Samuel and Tegmark, Max , title =. The First Conference on Language Modeling , year =
[22]

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

Meng, Zaiqiao and Liu, Fangyu and Shareghi, Ehsan and Su, Yixuan and Collins, Charlotte and Collier, Nigel , title =. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =
[23]

Advances in Neural Information Processing Systems , year =

Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , title =. Advances in Neural Information Processing Systems , year =
[24]

Journal of the Royal Statistical Society Series B: Statistical Methodology , volume =

Peters, Jonas and B. Causal inference by using invariant prediction: identification and confidence intervals , journal =. 2016 , month =. doi:10.1111/rssb.12167 , issn =

work page doi:10.1111/rssb.12167 2016
[25]

In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)

Petroni, Fabio and Rockt. Language models as knowledge bases? , booktitle =. 2019 , pages =. doi:10.18653/v1/D19-1250 , address =

work page doi:10.18653/v1/d19-1250 2019
[26]

Findings of the Association for Computational Linguistics: NAACL 2024 , year =

Pezeshkpour, Pouya and Hruschka, Estevam , title =. Findings of the Association for Computational Linguistics: NAACL 2024 , year =

2024
[27]

and Duvenaud, David , title =

Requeima, James and Bronskill, John and Choi, Dami and Turner, Richard E. and Duvenaud, David , title =. Advances in Neural Information Processing Systems , year =
[28]

and Cheng, Newton and Durmus, Esin and

Sharma, Mrinank and Tong, Meg and Korbak, Tomasz and Duvenaud, David and Askell, Amanda and Bowman, Samuel R. and Cheng, Newton and Durmus, Esin and. Towards understanding sycophancy in language models , booktitle =. 2024 , url =

2024
[29]

Logan IV, Eric Wallace, and Sameer Singh

Shin, Taylor and Razeghi, Yasaman and Logan Iv, Robert L. and Wallace, Eric and Singh, Sameer , title =. Proceedings of the 2020. 2020 , pages =. doi:10.18653/v1/2020.emnlp-main.346 , address =

work page doi:10.18653/v1/2020.emnlp-main.346 2020
[30]

Proceedings of the 2021

Sung, Mujeen and Lee, Jinhyuk and Yi, Sean and Jeon, Minji and Kim, Sungdong and Kang, Jaewoo , title =. Proceedings of the 2021. 2021 , pages =. doi:10.18653/v1/2021.emnlp-main.388 , address =

work page doi:10.18653/v1/2021.emnlp-main.388 2021
[31]

Transactions on Machine Learning Research , year =

Takayama, Masayuki and Okuda, Tadahisa and Pham, Thong and Ikenoue, Tatsuyoshi and Fukuma, Shingo and Shimizu, Shohei and Sannai, Akiyoshi , title =. Transactions on Machine Learning Research , year =
[32]

Thurstone, L. L. , title =. Psychological Review , year =. doi:10.1037/h0070288 , issn =

work page doi:10.1037/h0070288
[33]

Insights into ordinal embedding algorithms: a systematic evaluation , journal =

Vankadara, Leena Chennuru and Lohaus, Michael and Haghiri, Siavash and Wahab, Faiz Ul and. Insights into ordinal embedding algorithms: a systematic evaluation , journal =. 2023 , volume =

2023
[34]

and Sharma, Amit , title =

Vashishtha, Aniket and Reddy, Abbavaram Gowtham and Kumar, Abhinav and Bachu, Saketh and Balasubramanian, Vineeth N. and Sharma, Amit , title =. The Thirteenth International Conference on Learning Representations , year =
[35]

Advances in

Vig, Jesse and Gehrmann, Sebastian and Belinkov, Yonatan and Qian, Sharon and Nevo, Daniel and Singer, Yaron and Shieber, Stuart , title =. Advances in. 2020 , volume =

2020
[36]

Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , year =

Wan, Guangya and Lu, Yunsheng and Wu, Yuqi and Hu, Mengxuan and Li, Sheng , title =. Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , year =
[37]

2023 , journal =

Yao, Zonghai and Cao, Yi and Yang, Zhichao and Yu, Hong , title =. 2023 , journal =

2023
[38]

Give me the facts!

Youssef, Paul and Kora. Give me the facts!. Findings of the Association for Computational Linguistics: EMNLP 2023 , year =

2023
[39]

Causal parrots: large language models may talk causality but are not causal , journal =

Ze. Causal parrots: large language models may talk causality but are not causal , journal =. 2023 , url =

2023
[40]

2024 , doi =

Zhou, Yu and Wu, Xingyu and Huang, Beicheng and Wu, Jibin and Feng, Liang and Tan, Kay Chen , title =. 2024 , doi =

2024
[41]

Brodeur and Thomas A

Brodeur, Peter G. and Buckley, Thomas A. and Kanjee, Zahir and Goh, Ethan and Bin Ling, Evelyn and Jain, Priyank and Cabral, Stephanie and Abdulnour, Raja-Elie and Haimovich, Adrian D. and Freed, Jason A. and Olson, Andrew and Morgan, Daniel J. and Hom, Jason and Gallo, Robert and McCoy, Liam G. and Mombini, Haadi and Lucas, Christopher and Fotoohi, Misha...

work page doi:10.1126/science.adz4433
[42]

and Luce, R

Suppes, Patrick and Krantz, David M. and Luce, R. Duncan and Tversky, Amos , title =. 1989 , isbn =. doi:10.1016/C2009-0-21665-5 , url =

work page doi:10.1016/c2009-0-21665-5 1989
[43]

Prediction of air trapping or pulmonary hyperinflation by forced spirometry in

Alter, Peter and Orszag, Jan and Kellerer, Christoph and Kahnert, Kathrin and Speicher, Tobias and Watz, Henrik and Bals, Robert and Welte, Tobias and Vogelmeier, Claus and J. Prediction of air trapping or pulmonary hyperinflation by forced spirometry in. ERJ Open Research , year =. doi:10.1183/23120541.00092-2020 , url =

work page doi:10.1183/23120541.00092-2020 2020
[44]

and Annareddy, Srinivasulareddy , title =

Devalla, Lokesh and Ghewade, Bhalchandra and Jadhav, Ulhas S. and Annareddy, Srinivasulareddy , title =. Cureus , year =. doi:10.7759/cureus.53492 , url =

work page doi:10.7759/cureus.53492
[45]

PLOS ONE , year =

Holtz, Mayara and Perossi, Laura and Perossi, Juliana and dos Santos, Daniele Oliveira and de Souza, Hugo Celso Dutra and Gastaldi, Ada Clarice , title =. PLOS ONE , year =. doi:10.1371/journal.pone.0281780 , url =

work page doi:10.1371/journal.pone.0281780
[46]

and Bowen, Tomorrow and Luthra, Munish and Dixit, Adviteeya N

Hwang, Jiyoung and Saikawa, Eri and Avramov, Alexander and Chen, Siran and Shelly, Sandeep and Altartoor, Khaled A. and Bowen, Tomorrow and Luthra, Munish and Dixit, Adviteeya N. and Klein, Adam M. , title =. Environment International , year =. doi:10.1016/j.envint.2025.109560 , url =

work page doi:10.1016/j.envint.2025.109560 2025
[47]

Journal of Neuroradiology , year =

Nabizadeh, Farzaneh and Zafari, Roghayeh and Mohamadi, Mobin and Maleki, Tahereh and Fallahi, Mahdieh and Rafiei, Nazanin , title =. Journal of Neuroradiology , year =. doi:10.1016/j.neurad.2023.11.007 , url =

work page doi:10.1016/j.neurad.2023.11.007 2023
[48]

and Caba, Bastien and Elliott, Claire and Fisher, Elizabeth and de Moor, Chris and Belachew, Samia and Gafson, Arie , title =

Jiang, Xiaotong and Shen, Chunhua and Teunissen, Charlotte and Wessels, Marianne and Zetterberg, Henrik and Giovannoni, Gavin and Singh, Carol M. and Caba, Bastien and Elliott, Claire and Fisher, Elizabeth and de Moor, Chris and Belachew, Samia and Gafson, Arie , title =. Multiple Sclerosis Journal , year =. doi:10.1177/13524585231176732 , url =

work page doi:10.1177/13524585231176732
[49]

PLOS ONE , year =

Mirmosayyeb, Omid and Yazdan Panah, Mohammad and Mokary, Yousef and Mohammadi, Mohammad and Moases Ghaffary, Elham and Shaygannejad, Vahid and Weinstock-Guttman, Bianca and Zivadinov, Robert and Jakimovski, Dejan , title =. PLOS ONE , year =. doi:10.1371/journal.pone.0312421 , url =

work page doi:10.1371/journal.pone.0312421
[50]

Respiration , year =

Peng, Junjie and Li, Xiaohua and Zhou, Hong and Wang, Tao and Li, Xiaoou and Chen, Lei , title =. Respiration , year =. doi:10.1159/000541633 , url =

work page doi:10.1159/000541633
[51]

European Respiratory Review , year =

Wang, Tinggui and Liang, Mengdan and Ye, Xiaoying and Huang, Xiaoxiao and Chen, Hanbing and He, Xiongkun and Xie, Mengying and Xie, Xiaowei and Jiang, Xiannuan and Chen, Zhehui and Xie, Baosong and Zeng, Yiming and Xie, Xiaoxu , title =. European Respiratory Review , year =. doi:10.1183/16000617.0055-2025 , url =

work page doi:10.1183/16000617.0055-2025 2025
[52]

Association of serum glial fibrillary acidic protein with progression independent of relapse activity in multiple sclerosis , journal =

Rosenstein, Igal and Nordin, Anna and Sabir, Hemin and Malmestr. Association of serum glial fibrillary acidic protein with progression independent of relapse activity in multiple sclerosis , journal =. 2024 , volume =. doi:10.1007/s00415-024-12389-y , url =

work page doi:10.1007/s00415-024-12389-y 2024
[53]

and Everest, Marina and Beniameen, David and Rosehart, Heather , title =

Morrow, Sarah A. and Everest, Marina and Beniameen, David and Rosehart, Heather , title =. Multiple Sclerosis and Related Disorders , year =. doi:10.1016/j.msard.2025.106799 , url =

work page doi:10.1016/j.msard.2025.106799 2025
[54]

PLOS ONE , year =

Mollison, Daisy and Sellar, Robin and Bastin, Mark and Mollison, Denis and Chandran, Siddharthan and Wardlaw, Joanna and Connick, Peter , title =. PLOS ONE , year =. doi:10.1371/journal.pone.0177727 , url =

work page doi:10.1371/journal.pone.0177727
[55]

Seabold, J

Seabold, Skipper and Perktold, Josef , title =. Proceedings of the 9th Python in Science Conference , year =. doi:10.25080/Majora-92bf1922-011 , url =

work page doi:10.25080/majora-92bf1922-011
[56]

Introducing gpt-oss , year =