Elamur Research⌕
☰
home / papers / 2026 / Reliable Automated Triage in Spanish Clinical Notes: A Hybrid Framework for Risk-Aware HIV Suspicion Identification

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

Paper · 2026 · Rodrigo Morales-Sánchez, Soto Montalvo and Raquel Martínez

Summary

Clinical NLP classifiers that must assign a hard label to every case tend to be overconfident on ambiguous or atypical records, which is unsafe in medical triage. The authors address this for early detection of documented HIV suspicion in Spanish electronic health records. They introduce a post-hoc hybrid selective screening framework that can also defer a case to a human expert. It handles two kinds of uncertainty separately. Aleatoric ambiguity is managed with temperature-scaled Mondrian Conformal Prediction (MCP). Epistemic, out-of-distribution inputs are caught by a Multi-Centroid Mahalanobis Distance (MCMD) veto in a spectrally normalized latent space. A case is automated only if it passes both checks.

On a single-hospital cohort, conventional uncertainty methods lost most of their coverage under strict reliability targets. The hybrid policy kept high accuracy on a large automated share of patients. The deferral threshold works as an adjustable dial that trades safety against workload. The authors argue that this decoupling of uncertainty sources is needed before clinical NLP can be deployed responsibly.

Key points

  • Data: proprietary, anonymized EHRs from Hospital Universitario Fundación Alcorcón (HUFA, Madrid).
    • 13,642 patients, 63,802 notes, written 1998–2023.
    • 2,072 suspects, giving a class ratio of about 1:5.5.
    • Evaluation used 10-fold cross-validation plus a held-out test set.
  • Labels and leakage control:
    • Suspects are patients with a requested HIV serology, further filtered by EuroTEST indicator conditions.
    • Explicit HIV acronyms and testing recommendations were removed from the text to prevent leakage.
  • MCMD design:
    • The number of centroids per class is chosen by k-means with a 0.05 relative-inertia-gain cutoff.
    • All classes share one global precision matrix estimated with OAS.
    • Distance thresholds are class-specific, set at the 99th percentile.
  • Results:
    • At α = 0.01, Clear F2 is 0.982 with 67.7% of patients automated.
    • At α = 0.05, 91.3% are automated, Clear F2 is 0.913, and 10.2% of true positives are deferred.
    • Risk-aware metrics include Custom Risk-Kappa and True Positive Deferral Rate.
  • Interpretability and code:
    • Feature attribution shows that epistemic deferrals are driven by out-of-distribution terms such as psychiatric conditions.
    • Code is available at github.com/romorale/mil_uq_public.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Universidad Nacional de Educación a DistanciaMadrid · Spain6 works · 269 citations🎓Universidad Rey Juan CarlosMadrid · Spain4 works · 4 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksMedicine · 241 worksHealth Informatics · 184 worksHealth Professions · 69 worksHealth Information Management · 50 works

Cite

Cite as Rodrigo Morales-Sánchez, Soto Montalvo and Raquel Martínez. 2026. Reliable Automated Triage in Spanish Clinical Notes: A Hybrid Framework for Risk-Aware HIV Suspicion Identification. BioNLP 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.bionlp-1.8

@inproceedings{moralessanchez2026reliable,
  title = {{Reliable Automated Triage in Spanish Clinical Notes: A Hybrid Framework for Risk-Aware HIV Suspicion Identification}},
  author = {Rodrigo Morales-Sánchez, Soto Montalvo and Raquel Martínez},
  booktitle = {BioNLP 2026},
  year = {2026},
  doi = {10.18653/v1/2026.bionlp-1.8},
  url = {https://doi.org/10.18653/v1/2026.bionlp-1.8}
}

References

6 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Moloud Abdar, Farhad Pourpanah, Sadiq Hussain, Dana Rezazadegan, Li Liu, Mohammad Ghavamzadeh, Paul Fieguth, Xiaochun Cao, Abbas Khosravi, U. Rajendra Acharya, Vladimir Makarenkov, and Saeid Nahavandi. 2021. A review of uncertainty quantification in deep learning: Techniques, applications and challe… find
  2. Henrik Boström, Ulf Johansson, and Tuwe Löfström. 2021. Mondrian conformal predictive distributions. In Proceedings of the Tenth Symposium on Conformal and Probabilistic Prediction and Applications, volume 152 of Proceedings of Machine Learning Research, pages 24–38. PMLR. find
  3. Casimiro Pio Carrino, Joan Llop, Marc Pàmies, Asier Gutiérrez-Fandiño, Jordi Armengol-Estapé, Joaquín Silveira-Ocampo, Alfonso Valencia, Aitor Gonzalez-Agirre, and Marta Villegas. 2022. Pretrained Biomedical Language Models for Clinical NLP in Spanish. In Proceedings of the 21st Workshop on Biomedic… find
  4. EuroTEST. 2016. HIV Indicator Conditions: Guidance for Implementing HIV Testing in Adults in Health Care Settings. find
  5. Daniel J. Feller, Jason Zucker, Michael T. Yin, Peter Gordon, and Noémie Elhadad. 2018. Using Clinical Notes and Natural Language Processing for Automated HIV Risk Assessment. JAIDS Journal of Acquired Immune Deficiency Syndromes, 77(2):160–166. find
  6. Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: representing model uncertainty in deep learning. In Proceedings of the 33rd International Conference on Machine Learning, volume 48 of ICML’16, page 1050–1059. JMLR.org. \#\# Ethical Considerations find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sanchez-2026-reliable-automated-triage-spanish-clinical/