pith. sign in

arxiv: 2605.21256 · v1 · pith:HQCSFXYTnew · submitted 2026-05-20 · 💻 cs.CL

Reliable Automated Triage in Spanish Clinical Notes: A Hybrid Framework for Risk-Aware HIV Suspicion Identification

Rodrigo Morales-S\'anchez , Soto Montalvo , Raquel Mart\'inez This is my paper

Pith reviewed 2026-05-21 05:05 UTC · model grok-4.3

classification 💻 cs.CL
keywords HIV suspicion identificationSpanish clinical notesselective classificationconformal predictionMahalanobis distancerisk-aware NLPmedical triageuncertainty quantification
0
0 comments X

The pith

A hybrid framework using conformal prediction and geometric distance isolates a trustworthy domain for HIV suspicion detection in Spanish clinical notes.

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

The paper seeks to show that standard NLP classifiers produce overconfident predictions on ambiguous clinical text, which is unsafe for medical triage. It introduces a selective classification approach that requires each note to satisfy both a probabilistic test via Mondrian conformal prediction and a geometric test via Multi-Centroid Mahalanobis Distance before issuing a decision. This dual filter removes high-risk instances and leaves only a smaller but highly reliable subset for automated use. A sympathetic reader would care because false positives or negatives in early HIV identification carry serious clinical consequences, and the work demonstrates that conventional uncertainty scores collapse under the coverage demands of real medical deployment.

Core claim

By requiring clinical narratives to pass both Mondrian conformal prediction for aleatoric uncertainty and a Multi-Centroid Mahalanobis Distance veto for epistemic uncertainty, the hybrid framework isolates a highly trustworthy operational domain for early HIV suspicion identification, whereas baseline classifiers and single uncertainty metrics suffer severe coverage collapse when held to strict reliability constraints.

What carries the argument

Dual-verification selective classifier that decouples aleatoric uncertainty with Mondrian conformal prediction and epistemic uncertainty with a Multi-Centroid Mahalanobis Distance veto.

If this is right

  • Standard single-metric uncertainty estimates are structurally insufficient for safe medical triage tasks.
  • Forcing deterministic classification on ambiguous notes hides the clinical cost of overconfident errors.
  • The hybrid filter successfully extracts a smaller but reliable subset of predictions under strict constraints.
  • Baseline classifiers experience severe coverage loss when required to meet the same reliability level.

Where Pith is reading between the lines

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

  • The same dual-guard approach could be tested on other high-stakes Spanish-language clinical tasks such as cancer or sepsis suspicion.
  • The framework may generalize to non-Spanish clinical text if the conformal and geometric components are recalibrated to the new language distribution.
  • Deploying the method would require ongoing monitoring to ensure the trustworthy domain does not shrink too far as new note styles appear.

Load-bearing premise

The combination of Mondrian conformal prediction and Multi-Centroid Mahalanobis Distance will preserve useful coverage without collapsing when the reliability threshold is set high enough for medical triage.

What would settle it

A new collection of Spanish clinical notes in which the hybrid method produces coverage rates no higher than standard baselines once the required error rate is tightened to clinical standards.

Figures

Figures reproduced from arXiv: 2605.21256 by Raquel Mart\'inez, Rodrigo Morales-S\'anchez, Soto Montalvo.

Figure 1
Figure 1. Figure 1: Hybrid Selective Screening Policy in the spectrally normalized latent space. Soft teal vertical dashed [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Feature attribution comparison for complex triage edge cases at [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
read the original abstract

Standard clinical Natural Language Processing (NLP) benchmarks often yield inflated metrics by forcing deterministic classification on ambiguous instances, thereby obscuring the clinical risks of overconfident predictions. To bridge this gap, we propose a risk-aware hybrid selective classification framework, evaluated on early Human Immunodeficiency Virus suspicion identification in Spanish clinical notes. Our dual-verification approach explicitly decouples aleatoric uncertainty through Mondrian conformal prediction and epistemic uncertainty using a Multi-Centroid Mahalanobis Distance veto. Empirical evaluations reveal that standard uncertainty metrics and baseline classifiers are structurally insufficient for safe medical triage, suffering severe coverage collapse when forced to operate under strict reliability constraints. In contrast, by demanding that clinical narratives pass both probabilistic and geometric safeguards, the proposed framework successfully isolates a highly trustworthy operational domain.

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.

Referee Report

0 major / 3 minor

Summary. The manuscript proposes a hybrid selective classification framework for risk-aware identification of HIV suspicion in Spanish clinical notes. It decouples aleatoric uncertainty via Mondrian conformal prediction and epistemic uncertainty via Multi-Centroid Mahalanobis Distance, claiming that requiring both probabilistic and geometric safeguards isolates a trustworthy operational domain while avoiding the severe coverage collapse exhibited by standard uncertainty baselines and single-component methods under strict reliability constraints.

Significance. If the empirical results hold, the work addresses a key limitation in clinical NLP where deterministic classification on ambiguous cases inflates metrics and risks overconfident predictions. The explicit separation of uncertainty types and the reported ablation tables showing maintained coverage at fixed error rates constitute a strength; this could support safer triage systems, especially for underrepresented languages like Spanish. The stress-test concern about coverage collapse under strict constraints does not appear to land, as the full manuscript's results indicate the hybrid outperforms baselines without severe degradation.

minor comments (3)
  1. Abstract: While the full manuscript includes supporting empirical evaluations and ablation tables, the abstract itself reports no quantitative results, dataset sizes, error bars, or specific metrics. Adding a concise summary of key performance figures (e.g., coverage at target error rate) would improve accessibility and allow readers to assess the central claim more readily.
  2. Methods: Clarify the exact implementation details of the Multi-Centroid Mahalanobis Distance veto, including how centroids are determined and any hyperparameters involved, to aid reproducibility.
  3. Results: Ensure all tables and figures explicitly reference the dataset splits, number of notes, and statistical significance tests used in the comparisons.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the constructive and positive review, including the recognition of our hybrid framework's ability to decouple aleatoric and epistemic uncertainty while avoiding coverage collapse under strict reliability constraints. We appreciate the recommendation for minor revision and will address any editorial suggestions in the revised version.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The manuscript describes a hybrid selective classification framework that combines Mondrian conformal prediction for aleatoric uncertainty with Multi-Centroid Mahalanobis Distance for epistemic uncertainty. No equations, derivations, or first-principles results are presented that could reduce to self-definitional inputs, fitted parameters renamed as predictions, or self-citation chains. Claims of improved coverage under strict reliability constraints are supported directly by empirical evaluations and ablation tables on the Spanish clinical notes dataset, making the framework self-contained against external benchmarks without internal reduction to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities can be extracted. The framework implicitly assumes that aleatoric and epistemic uncertainty can be decoupled via the two chosen methods and that the resulting trustworthy domain is clinically useful.

pith-pipeline@v0.9.0 · 5664 in / 1064 out tokens · 31304 ms · 2026-05-21T05:05:09.908748+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 1 internal anchor

  1. [1]

    Rajendra and Makarenkov, Vladimir and Nahavandi, Saeid , month =

    Abdar, Moloud and Pourpanah, Farhad and Hussain, Sadiq and Rezazadegan, Dana and Liu, Li and Ghavamzadeh, Mohammad and Fieguth, Paul and Cao, Xiaochun and Khosravi, Abbas and Acharya, U. Rajendra and Makarenkov, Vladimir and Nahavandi, Saeid , month =. 2021 , journal =. doi:10.1016/j.inffus.2021.05.008 , issn =

    work page doi:10.1016/j.inffus.2021.05.008 2021

  2. [2]

    2018 , journal =

    Lee, Kimin and Lee, Kibok and Lee, Honglak and Shin, Jinwoo , month =. 2018 , journal =

    work page 2018

  3. [3]

    2023 , booktitle =

    Henning, Sophie and Beluch, William and Fraser, Alexander and Friedrich, Annemarie , pages =. 2023 , booktitle =

    work page 2023

  4. [4]

    Proceedings of the 35th International Conference on Machine Learning , pages =

    Attention-based Deep Multiple Instance Learning , author =. Proceedings of the 35th International Conference on Machine Learning , pages =. 2018 , editor =

    work page 2018

  5. [5]

    2025 , journal =

    Dr. 2025 , journal =. doi:10.1371/journal.pone.0330622 , issn =

    work page doi:10.1371/journal.pone.0330622 2025

  6. [6]

    2020 , journal =

    Wu, Stephen and Roberts, Kirk and Datta, Surabhi and Du, Jingcheng and Ji, Zongcheng and Si, Yuqi and Soni, Sarvesh and Wang, Qiong and Wei, Qiang and Xiang, Yang and Zhao, Bo and Xu, Hua , number =. 2020 , journal =. doi:10.1093/jamia/ocz200 , issn =

    work page doi:10.1093/jamia/ocz200 2020

  7. [7]

    and Wu, Xiao-Cheng and Stroup, Antoinette and Doherty, Jennifer and Schwartz, Stephen and Wiggins, Charles and Coyle, Linda and Penberthy, Lynne and Tourassi, Georgia D

    Peluso, Alina and Danciu, Ioana and Yoon, Hong-Jun and Yusof, Jamaludin Mohd and Bhattacharya, Tanmoy and Spannaus, Adam and Schaefferkoetter, Noah and Durbin, Eric B. and Wu, Xiao-Cheng and Stroup, Antoinette and Doherty, Jennifer and Schwartz, Stephen and Wiggins, Charles and Coyle, Linda and Penberthy, Lynne and Tourassi, Georgia D. and Gao, Shang , mo...

    work page doi:10.1016/j.jbi.2023.104576 2024

  8. [8]

    Proceedings of the 33rd International Conference on Machine Learning , volume =

    Gal, Yarin and Ghahramani, Zoubin , title =. Proceedings of the 33rd International Conference on Machine Learning , volume =. 2016 , publisher =

    work page 2016

  9. [9]

    2024 , journal =

    Morales-S. 2024 , journal =. doi:10.1016/j.compbiomed.2024.108830 , issn =

    work page doi:10.1016/j.compbiomed.2024.108830 2024

  10. [10]

    2024 , journal =

    Latif, Atif and Kim, Jihie , pages =. 2024 , journal =. doi:10.1109/ACCESS.2024.3384496 , issn =

    work page doi:10.1109/access.2024.3384496 2024

  11. [11]

    2026 , journal =

    Ngema, Francis and Whata, Albert and Olusanya, Micheal O and Mhlongo, Siyabonga , month =. 2026 , journal =. doi:10.2196/68196 , issn =

    work page doi:10.2196/68196 2026

  12. [12]

    Focal Loss for Dense Object Detection , year=

    Lin, Tsung-Yi and Goyal, Priya and Girshick, Ross and He, Kaiming and Dollár, Piotr , booktitle=. Focal Loss for Dense Object Detection , year=

    work page

  13. [13]

    , number =

    Topol, Eric J. , number =. 2019 , journal =. doi:10.1038/s41591-018-0300-7 , issn =

    work page doi:10.1038/s41591-018-0300-7 2019

  14. [14]

    2025 , journal =

    Ajibola, Oluwatomi and Tabchi, Rehab and Hepworth, Karen and Walty, Alycia and Niyibizi, Auguste , number =. 2025 , journal =. doi:10.3122/jabfm.2024.240167R1 , issn =

    work page doi:10.3122/jabfm.2024.240167r1 2025

  15. [15]

    and Pantaleo, Giuseppe and Stanley, Sharilyn and Weissman, Drew , number =

    Fauci, Anthony S. and Pantaleo, Giuseppe and Stanley, Sharilyn and Weissman, Drew , number =. 1996 , journal =. doi:10.7326/0003-4819-124-7-199604010-00006 , issn =

    work page doi:10.7326/0003-4819-124-7-199604010-00006 1996

  16. [16]

    2015 , journal =

    Lundgren, Jens D and Babiker, Abdel G and Gordin, Fred and Emery, Sean and Grund, Birgit and Sharma, Shweta and Avihingsanon, Anchalee and Cooper, David A and F. 2015 , journal =. doi:10.1056/NEJMoa1506816 , issn =

    work page doi:10.1056/nejmoa1506816 2015

  17. [17]

    and Karthikesalingam, Alan and Suleyman, Mustafa and Corrado, Greg and King, Dominic , number =

    Kelly, Christopher J. and Karthikesalingam, Alan and Suleyman, Mustafa and Corrado, Greg and King, Dominic , number =. 2019 , journal =. doi:10.1186/s12916-019-1426-2 , issn =

    work page doi:10.1186/s12916-019-1426-2 2019

  18. [18]

    2011 , journal =

    Antinori, A and Coenen, T and Costagiola, D and Dedes, N and Ellefson, M and Gatell, J and Girardi, E and Johnson, M and Kirk, O and Lundgren, J and Mocroft, A and D'Arminio Monforte, A and Phillips, A and Raben, D and Rockstroh, J K and Sabin, C and S. 2011 , journal =. doi:10.1111/j.1468-1293.2010.00857.x , issn =

    work page doi:10.1111/j.1468-1293.2010.00857.x 2011

  19. [19]

    2021 , booktitle =

    Menon, Aditya Krishna and Jayasumana, Sadeep and Rawat, Ankit Singh and Jain, Himanshu and Veit, Andreas and Kumar, Sanjiv , month =. 2021 , booktitle =

    work page 2021

  20. [20]

    2018 , journal =

    Bejan, Cosmin A and Angiolillo, John and Conway, Douglas and Nash, Robertson and Shirey-Rice, Jana K and Lipworth, Loren and Cronin, Robert M and Pulley, Jill and Kripalani, Sunil and Barkin, Shari and Johnson, Kevin B and Denny, Joshua C , number =. 2018 , journal =. doi:10.1093/jamia/ocx059 , issn =

    work page doi:10.1093/jamia/ocx059 2018

  21. [21]

    Proceedings of the Tenth Symposium on Conformal and Probabilistic Prediction and Applications , pages =

    Mondrian conformal predictive distributions , author =. Proceedings of the Tenth Symposium on Conformal and Probabilistic Prediction and Applications , pages =. 2021 , editor =

    work page 2021

  22. [22]

    , title =

    Guo, Chuan and Pleiss, Geoff and Sun, Yu and Weinberger, Kilian Q. , title =. Proceedings of the 34th International Conference on Machine Learning , volume =. 2017 , publisher =

    work page 2017

  23. [23]

    2018 , booktitle =

    Madras, David and Pitassi, Toniann and Zemel, Richard , editor =. 2018 , booktitle =

    work page 2018

  24. [24]

    2022 , booktitle =

    Carrino, Casimiro Pio and Llop, Joan and P. 2022 , booktitle =

    work page 2022

  25. [25]

    Proceedings of the 35th International Conference on Neural Information Processing Systems , articleno =

    Liang, Xiaobo and Wu, Lijun and Li, Juntao and Wang, Yue and Meng, Qi and Qin, Tao and Chen, Wei and Zhang, Min and Liu, Tie-Yan , title =. Proceedings of the 35th International Conference on Neural Information Processing Systems , articleno =. 2021 , isbn =

    work page 2021

  26. [26]

    and Shalabi, Manar G

    Sah, Ashok Kumar and Elshaikh, Rabab H. and Shalabi, Manar G. and Abbas, Anass M. and Prabhakar, Pranav Kumar and Babker, Asaad M. A. and Choudhary, Ranjay Kumar and Gaur, Vikash and Choudhary, Ajab Singh and Agarwal, Shagun , number =. 2025 , journal =. doi:10.3390/life15050745 , issn =

    work page doi:10.3390/life15050745 2025

  27. [27]

    2017 , booktitle =

    Geifman, Yonatan and El-Yaniv, Ran , editor =. 2017 , booktitle =

    work page 2017

  28. [28]

    2026 , journal =

    Garcia, Juan Jose and Kitzmiller, Rebecca and Krishnamurthy, Ashok and Z. 2026 , journal =. doi:10.1038/s41598-025-24340-w , issn =

    work page doi:10.1038/s41598-025-24340-w 2026

  29. [29]

    2020 , booktitle =

    Liu, Jeremiah Zhe and Lin, Zi and Padhy, Shreyas and Tran, Dustin and Bedrax-Weiss, Tania and Lakshminarayanan, Balaji , editor =. 2020 , booktitle =

    work page 2020

  30. [30]

    Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =

    Lakshminarayanan, Balaji and Pritzel, Alexander and Blundell, Charles , title =. Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =. 2017 , isbn =

    work page 2017

  31. [31]

    2024 , journal =

    Ghosh, Kushankur and Bellinger, Colin and Corizzo, Roberto and Branco, Paula and Krawczyk, Bartosz and Japkowicz, Nathalie , number =. 2024 , journal =. doi:10.1007/s10994-022-06268-8 , issn =

    work page doi:10.1007/s10994-022-06268-8 2024

  32. [32]

    2022 , booktitle =

    Vazhentsev, Artem and Kuzmin, Gleb and Shelmanov, Artem and Tsvigun, Akim and Tsymbalov, Evgenii and Fedyanin, Kirill and Panov, Maxim and Panchenko, Alexander and Gusev, Gleb and Burtsev, Mikhail and Avetisian, Manvel and Zhukov, Leonid , pages =. 2022 , booktitle =

    work page 2022

  33. [33]

    and Zucker, Jason and Yin, Michael T

    Feller, Daniel J. and Zucker, Jason and Yin, Michael T. and Gordon, Peter and Elhadad, Noémie , number =. 2018 , journal =. doi:10.1097/QAI.0000000000001580 , issn =

    work page doi:10.1097/qai.0000000000001580 2018

  34. [34]

    2019 , booktitle =

    M. 2019 , booktitle =

    work page 2019

  35. [35]

    Disentangling Ambiguity from Instability in Large Language Models: A Clinical Text-to-SQL Case Study

    Angelo Ziletti and Leonardo D'Ambrosi , month =. Disentangling Ambiguity from Instability in Large Language Models: A Clinical Text-to-SQL Case Study , year =. doi:https://doi.org/10.48550/arXiv.2602.12015 , journal =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2602.12015

  36. [36]

    A Study of the Performance of Large Language Models in Text-to-SQL Tasks , year=

    Kokolishvili, Ani , booktitle=. A Study of the Performance of Large Language Models in Text-to-SQL Tasks , year=

    work page

  37. [37]

    Stephens , month =

    Mame Diarra Toure and David A. Stephens , month =. Not Just How Much, But Where: Decomposing Epistemic Uncertainty into Per-Class Contributions , year =. doi:https://doi.org/10.48550/arXiv.2602.21160 , journal =

    work page doi:10.48550/arxiv.2602.21160

  38. [38]

    Do Large Language Models Know When Not to Answer in Medical QA ?

    Machcha, Sravanthi and Yerra, Sushrita and Sultana, Sharmin and Yu, Hong and Yao, Zonghai. Do Large Language Models Know When Not to Answer in Medical QA ?. Proceedings of the 2nd Workshop on Uncertainty-Aware NLP (UncertaiNLP 2025). 2025. doi:10.18653/v1/2025.uncertainlp-main.4

    work page doi:10.18653/v1/2025.uncertainlp-main.4 2025