🩺 Medicine

A survey · about 4 minutes to read · every paper named here has its own page in this library

The Medicine shelf collects the papers of this library that OpenAlex places in medicine, mostly language processing built for biology and health: finding named entities in research abstracts, adapting pretrained models to clinical notes, summarising patient records, simplifying medical texts for lay readers, and structuring radiology reports. It is a small corner of this library, and most of its papers also stand on the Artificial Intelligence shelf.

What the shelf does show is the arc of biomedical NLP in miniature: a classic shared task from the early 2000s, a careful benchmark of domain-adapted transformers from 2020, and a cluster of 2026 workshop papers that apply large and small language models to clinical text in French, Hebrew, German and English.

Scope

Biomedical text differs from news or web text in vocabulary, abbreviation habits, naming conventions and, in the clinic, in privacy constraints and terse note-taking. Research in the area has two broad branches: work on the scientific literature (abstracts and full texts in PubMed) and work on clinical documents (patient records, discharge letters, imaging reports). Both branches appear here, alongside one paper on health-related social-media-style classification of research sentences.

History in brief

Biomedical named entity recognition became a community benchmark task in the early 2000s, built on the GENIA corpus of annotated MEDLINE abstracts. The late 2010s brought domain-specific pretrained encoders such as BioBERT, SciBERT and ClinicalBERT, and with them the question of whether domain pretraining beats a stronger general model. The most recent papers move to retrieval-augmented generation, small open models and uncertainty estimation, and to languages other than English.

Papers in this library

A benchmark for biomedical entity recognition. Introduction to the Bio-entity Recognition Task at JNLPBA (2004) describes the shared task that asked systems to tag proteins, DNA, RNA, cell types and cell lines in MEDLINE abstracts. The organisers collapsed GENIA's fine-grained classes into five, released data in IOB2 format, and split the test abstracts by publication period to probe generalisation. Most participants used support vector machines, often combined with sequence models; even the best total F-score stayed under 73, well behind newswire systems of the time. The data set remained a standard reference for biomedical NER and is still among the tasks used to test later models.

Pretrained models for biomedical and clinical text. Pretrained Language Models for Biomedical and Clinical Tasks: Understanding and Extending the State-of-the-Art (2020) put existing domain models and general RoBERTa checkpoints through one fine-tuning protocol on 18 established tasks, JNLPBA among them. Domain models led on literature tasks, while the large general model did better on clinical ones. The authors then pretrained RoBERTa-style models on freely available PubMed, PubMed Central and MIMIC-III text; a large model with an in-domain vocabulary and mixed data set new best results on half the tasks, and the models were released.

Retrieval for French patient records. Evaluating the Retrieval Component in a Retrieval-Augmented Summarization System for Patient Records in French (2026) studies the first stage of a summariser for intensive-care clinicians. Queries built from physicians' annotations retrieved relevant chunks better than hand-written prompts, especially with simple TF-IDF; see also information retrieval & search and summarization.

Simplifying mental-health definitions. Evaluation of Multilingual Text Simplification for the Mental Health Domain: Exploring Small Language Models (2026) tried nine small models on ICD-11 articles about psychotic disorders in three languages. Only one model gave usable output, and only in English; changes in lexical diversity and readability proved more telling than similarity scores.

Classifying sentences about patients. PEI at #SMM4H-HeaRD 2026: Enhancing Patient Metadata Detection via Hypothesis-Conditioned Classification and Paraphrase-Based Data Augmentation (2026) found that a biomedical BERT encoder, recast as a hypothesis-pairing task and trained on paraphrase-augmented data, beat a fine-tuned generative model at spotting patient details in COVID-19 sequencing papers.

Uncertainty in Hebrew radiology. Trustworthy NLP for Low-Resource Languages: Agent-Based Uncertainty Modeling for Hebrew Radiology Report Structuring (2026) asks a model the same question through several paraphrased prompts and lets a larger model grade its confidence; withholding uncertain cases raised accuracy at the cost of coverage. It links the shelf to Hebrew & Aramaic and low-resource languages.

People and places

The JNLPBA task was organised by Nigel Collier, now associated with the University of Cambridge and University College London, together with Jin-Dong Kim, Tomoko Ohta and Yoshimasa Tsuruoka, names closely tied to the GENIA project at The University of Tokyo. The 2020 benchmark came from Patrick Lewis, Myle Ott, Jingfei Du and Veselin Stoyanov; Lewis is now at Cohere and Stoyanov at Mohamed bin Zayed University of Artificial Intelligence.

French clinical NLP is represented by Aurélie Névéol of the Centre National de la Recherche Scientifique and Xavier Tannier, linked with Inserm and Sorbonne Université. The Hebrew radiology work joins engineers such as Moti Freiman and Hadas Ben Atya of the Technion with clinicians including Dan Turner of the Hebrew University of Jerusalem.

Open questions

The papers here point to several unsolved issues: how to evaluate simplification and summarisation of medical text without costly expert judgement; whether small, locally run models can be trusted with clinical documents in languages with little training data; and when a system should abstain rather than answer. The gap between English and other languages is visible throughout.

Neighbouring shelves

All of the shelf's papers also sit on Artificial Intelligence; some cross into Language and Linguistics and Machine translation through their multilingual work. Related topics include evaluation & benchmarks, language models & pretraining and multilingual & cross-lingual NLP.

The library in figures