Xavier Tannier

Christel Gérardin, Yuhan Xiong, Perceval Wajsbürt et al.

The objective of our study is to determine whether using English tools to extract and normalize French medical concepts on translations provides comparable performance to French models trained on a set of annotated French clinical notes. We compare two methods: a method involving French language models and a method involving English language models. For the native French method, the Named Entity Recognition (NER) and normalization steps are performed separately. For the translated English method, after the first translation step, we compare a two-step method and a terminology-oriented method that performs extraction and normalization at the same time. We used French, English and bilingual annotated datasets to evaluate all steps (NER, normalization and translation) of our algorithms. Concerning the results, the native French method performs better than the translated English one with a global f1 score of 0.51 [0.47;0.55] against 0.39 [0.34;0.44] and 0.38 [0.36;0.40] for the two English methods tested. In conclusion, despite the recent improvement of the translation models, there is a significant performance difference between the two approaches in favor of the native French method which is more efficient on French medical texts, even with few annotated documents.

Xavier Tannier, Perceval Wajsbürt, Alice Calliger et al.

The objective of this study is to address the critical issue of de-identification of clinical reports in order to allow access to data for research purposes, while ensuring patient privacy. The study highlights the difficulties faced in sharing tools and resources in this domain and presents the experience of the Greater Paris University Hospitals (AP-HP) in implementing a systematic pseudonymization of text documents from its Clinical Data Warehouse. We annotated a corpus of clinical documents according to 12 types of identifying entities, and built a hybrid system, merging the results of a deep learning model as well as manual rules. Our results show an overall performance of 0.99 of F1-score. We discuss implementation choices and present experiments to better understand the effort involved in such a task, including dataset size, document types, language models, or rule addition. We share guidelines and code under a 3-Clause BSD license.

Basile Dura, Charline Jean, Xavier Tannier et al.

Background Clinical studies using real-world data may benefit from exploiting clinical reports, a particularly rich albeit unstructured medium. To that end, natural language processing can extract relevant information. Methods based on transfer learning using pre-trained language models have achieved state-of-the-art results in most NLP applications; however, publicly available models lack exposure to speciality-languages, especially in the medical field. Objective We aimed to evaluate the impact of adapting a language model to French clinical reports on downstream medical NLP tasks. Methods We leveraged a corpus of 21M clinical reports collected from August 2017 to July 2021 at the Greater Paris University Hospitals (APHP) to produce two CamemBERT architectures on speciality language: one retrained from scratch and the other using CamemBERT as its initialisation. We used two French annotated medical datasets to compare our language models to the original CamemBERT network, evaluating the statistical significance of improvement with the Wilcoxon test. Results Our models pretrained on clinical reports increased the average F1-score on APMed (an APHP-specific task) by 3 percentage points to 91%, a statistically significant improvement. They also achieved performance comparable to the original CamemBERT on QUAERO. These results hold true for the fine-tuned and from-scratch versions alike, starting from very few pre-training samples. Conclusions We confirm previous literature showing that adapting generalist pre-train language models such as CamenBERT on speciality corpora improves their performance for downstream clinical NLP tasks. Our results suggest that retraining from scratch does not induce a statistically significant performance gain compared to fine-tuning.

Xavier Tannier, Salam Abbara, Rémi Flicoteaux et al.

The development of clinical natural language processing (NLP) systems is severely hampered by the sensitive nature of medical records, which restricts data sharing under stringent privacy regulations, particularly in France and the broader European Union. To address this gap, we introduce PARHAF, a large open-source corpus of clinical documents in French. PARHAF comprises expert-authored clinical reports describing realistic yet entirely fictitious patient cases, making it anonymous and freely shareable by design. The corpus was developed using a structured protocol that combined clinician expertise with epidemiological guidance from the French National Health Data System (SNDS), ensuring broad clinical coverage. A total of 104 medical residents across 18 specialties authored and peer-reviewed the reports following predefined clinical scenarios and document templates. The corpus contains 7394 clinical reports covering 5009 patient cases across a wide range of medical and surgical specialties. It includes a general-purpose component designed to approximate real-world hospitalization distributions, and four specialized subsets that support information-extraction use cases in oncology, infectious diseases, and diagnostic coding. Documents are released under a CC-BY open license, with a portion temporarily embargoed to enable future benchmarking under controlled conditions. PARHAF provides a valuable resource for training and evaluating French clinical language models in a fully privacy-preserving setting, and establishes a replicable methodology for building shareable synthetic clinical corpora in other languages and health systems.

Aidan Mannion, Cécile Macaire, Armand Violle et al.

Large language models (LLMs) have demonstrated remarkable capabilities across diverse domains, yet their adaptation to specialized fields remains challenging, particularly for non-English languages. This study investigates domain-adaptive pre-training (DAPT) as a strategy for specializing small to mid-sized LLMs in the French biomedical domain through continued pre-training. We address two key research questions: the viability of specialized continued pre-training for domain adaptation and the relationship between domain-specific performance gains and general capability degradation. Our contributions include the release of a fully open-licensed French biomedical corpus suitable for commercial and open-source applications, the training and release of specialized French biomedical LLMs, and novel insights for DAPT implementation. Our methodology encompasses the collection and refinement of high-quality French biomedical texts, the exploration of causal language modeling approaches using DAPT, and conducting extensive comparative evaluations. Our results cast doubt on the efficacy of DAPT, in contrast to previous works, but we highlight its viability in smaller-scale, resource-constrained scenarios under the right conditions. Findings in this paper further suggest that model merging post-DAPT is essential to mitigate generalization trade-offs, and in some cases even improves performance on specialized tasks at which the DAPT was directed.

Adam Remaki, Christel Gérardin, Eulàlia Farré-Maduell et al.

We present SynCABEL (Synthetic Contextualized Augmentation for Biomedical Entity Linking), a framework that addresses a central bottleneck in supervised biomedical entity linking (BEL): the scarcity of expert-annotated training data. SynCABEL leverages large language models to generate context-rich synthetic training examples for all candidate concepts in a target knowledge base, providing broad supervision without manual annotation. We demonstrate that SynCABEL, when combined with decoder-only models and guided inference establish new state-of-the-art results across three widely used multilingual benchmarks: MedMentions for English, QUAERO for French, and SPACCC for Spanish. Evaluating data efficiency, we show that SynCABEL reaches the performance of full human supervision using up to 60% less annotated data, substantially reducing reliance on labor-intensive and costly expert labeling. Finally, acknowledging that standard evaluation based on exact code matching often underestimates clinically valid predictions due to ontology redundancy, we introduce an LLM-as-a-judge protocol. This analysis reveals that SynCABEL significantly improves the rate of clinically valid predictions. Our synthetic datasets, models, and code are released to support reproducibility and future research.

Adam Remaki, Xavier Tannier, Christel Gérardin

Biomedical entity linking maps textual mentions to concepts in structured knowledge bases such as UMLS or SNOMED CT. Most existing systems link each mention independently, using only the mention or its surrounding sentence. This ignores dependencies between mentions in the same document and can lead to inconsistent predictions, especially when the same concept appears under different surface forms. We introduce LongBEL, a document-level generative framework that combines full-document context with a memory of previous predictions. To make this memory robust, LongBEL is trained with cross-validated predictions rather than gold labels, reducing the mismatch between training and inference and limiting cascading errors. Experiments on five biomedical benchmarks across English, French, and Spanish show that LongBEL improves over sentence-level generative baselines, with the largest gains on datasets where concepts frequently recur within documents. An ensemble of local, global, and memory-based variants achieves the best results across all benchmarks. Further analysis shows that the largest gains occur on recurring concepts, suggesting that LongBEL mainly improves document-level consistency rather than isolated mention disambiguation.

Jamil Zaghir, Marco Naguib, Mina Bjelogrlic et al.

Prompt engineering is crucial for harnessing the potential of large language models (LLMs), especially in the medical domain where specialized terminology and phrasing is used. However, the efficacy of prompt engineering in the medical domain remains to be explored. In this work, 114 recent studies (2022-2024) applying prompt engineering in medicine, covering prompt learning (PL), prompt tuning (PT), and prompt design (PD) are reviewed. PD is the most prevalent (78 articles). In 12 papers, PD, PL, and PT terms were used interchangeably. ChatGPT is the most commonly used LLM, with seven papers using it for processing sensitive clinical data. Chain-of-Thought emerges as the most common prompt engineering technique. While PL and PT articles typically provide a baseline for evaluating prompt-based approaches, 64% of PD studies lack non-prompt-related baselines. We provide tables and figures summarizing existing work, and reporting recommendations to guide future research contributions.

Marco Naguib, Xavier Tannier, Aurélie Névéol

Large language models (LLMs) have become the preferred solution for many natural language processing tasks. In low-resource environments such as specialized domains, their few-shot capabilities are expected to deliver high performance. Named Entity Recognition (NER) is a critical task in information extraction that is not covered in recent LLM benchmarks. There is a need for better understanding the performance of LLMs for NER in a variety of settings including languages other than English. This study aims to evaluate generative LLMs, employed through prompt engineering, for few-shot clinical NER. %from the perspective of F1 performance and environmental impact. We compare 13 auto-regressive models using prompting and 16 masked models using fine-tuning on 14 NER datasets covering English, French and Spanish. While prompt-based auto-regressive models achieve competitive F1 for general NER, they are outperformed within the clinical domain by lighter biLSTM-CRF taggers based on masked models. Additionally, masked models exhibit lower environmental impact compared to auto-regressive models. Findings are consistent across the three languages studied, which suggests that LLM prompting is not yet suited for NER production in the clinical domain.

Chi-en Amy Tai, Xavier Tannier

Clinical trials are a critical process in the medical field for introducing new treatments and innovations. However, cohort selection for clinical trials is a time-consuming process that often requires manual review of patient text records for specific keywords. Though there have been studies on standardizing the information across the various platforms, Natural Language Processing (NLP) tools remain crucial for spotting eligibility criteria in textual reports. Recently, pre-trained large language models (LLMs) have gained popularity for various NLP tasks due to their ability to acquire a nuanced understanding of text. In this paper, we study the performance of large language models on clinical trial cohort selection and leverage the n2c2 challenges to benchmark their performance. Our results are promising with regard to the incorporation of LLMs for simple cohort selection tasks, but also highlight the difficulties encountered by these models as soon as fine-grained knowledge and reasoning are required.

Guillaume Bazin, Xavier Tannier, Fanny Adda et al.

Rules could be an information extraction (IE) default option, compared to ML and LLMs in terms of sustainability, transferability, interpretability, and development burden. We suggest a sustainable and combined use of rules and ML as an IE method. Our approach starts with an exhaustive expert manual highlighting in a single working session of a representative subset of the data corpus. We developed and validated the feasibility and the performance metrics of the REST decision tool to help the annotator choose between rules as a by default option and ML for each entity of an IE task. REST makes the annotator visualize the characteristics of each entity formalization in the free texts and the expected rule development feasibility and IE performance metrics. ML is considered as a backup IE option and manual annotation for training is therefore minimized. The external validity of REST on a 12-entity use case showed good reproducibility.

Nesrine Bannour, Christophe Servan, Aurélie Névéol et al.

Background: Transformer-based language models have shown strong performance on many Natural LanguageProcessing (NLP) tasks. Masked Language Models (MLMs) attract sustained interest because they can be adaptedto different languages and sub-domains through training or fine-tuning on specific corpora while remaining lighterthan modern Large Language Models (LLMs). Recently, several MLMs have been released for the biomedicaldomain in French, and experiments suggest that they outperform standard French counterparts. However, nosystematic evaluation comparing all models on the same corpora is available. Objective: This paper presentsan evaluation of masked language models for biomedical French on the task of clinical named entity recognition.Material and methods: We evaluate biomedical models CamemBERT-bio and DrBERT and compare them tostandard French models CamemBERT, FlauBERT and FrALBERT as well as multilingual mBERT using three publicallyavailable corpora for clinical named entity recognition in French. The evaluation set-up relies on gold-standardcorpora as released by the corpus developers. Results: Results suggest that CamemBERT-bio outperformsDrBERT consistently while FlauBERT offers competitive performance and FrAlBERT achieves the lowest carbonfootprint. Conclusion: This is the first benchmark evaluation of biomedical masked language models for Frenchclinical entity recognition that compares model performance consistently on nested entity recognition using metricscovering performance and environmental impact.

Christel Gérardin, Perceval Wajsbürt, Basile Dura et al.

Objective:Develop and validate an algorithm for analyzing the layout of PDF clinical documents to improve the performance of downstream natural language processing tasks. Materials and Methods: We designed an algorithm to process clinical PDF documents and extract only clinically relevant text. The algorithm consists of several steps: initial text extraction using a PDF parser, followed by classification into categories such as body text, left notes, and footers using a Transformer deep neural network architecture, and finally an aggregation step to compile the lines of a given label in the text. We evaluated the technical performance of the body text extraction algorithm by applying it to a random sample of documents that were annotated. Medical performance was evaluated by examining the extraction of medical concepts of interest from the text in their respective sections. Finally, we tested an end-to-end system on a medical use case of automatic detection of acute infection described in the hospital report. Results:Our algorithm achieved per-line precision, recall, and F1 score of 98.4, 97.0, and 97.7, respectively, for body line extraction. The precision, recall, and F1 score per document for the acute infection detection algorithm were 82.54 (95CI 72.86-91.60), 85.24 (95CI 76.61-93.70), 83.87 (95CI 76, 92-90.08) with exploitation of the results of the advanced body extraction algorithm, respectively. Conclusion:We have developed and validated a system for extracting body text from clinical documents in PDF format by identifying their layout. We were able to demonstrate that this preprocessing allowed us to obtain better performances for a common downstream task, i.e., the extraction of medical concepts in their respective sections, thus proving the interest of this method on a clinical use case.

Adrian Ahne, Vivek Khetan, Xavier Tannier et al.

Objective: Leveraging machine learning methods, we aim to extract both explicit and implicit cause-effect associations in patient-reported, diabetes-related tweets and provide a tool to better understand opinion, feelings and observations shared within the diabetes online community from a causality perspective. Materials and Methods: More than 30 million diabetes-related tweets in English were collected between April 2017 and January 2021. Deep learning and natural language processing methods were applied to focus on tweets with personal and emotional content. A cause-effect-tweet dataset was manually labeled and used to train 1) a fine-tuned Bertweet model to detect causal sentences containing a causal association 2) a CRF model with BERT based features to extract possible cause-effect associations. Causes and effects were clustered in a semi-supervised approach and visualised in an interactive cause-effect-network. Results: Causal sentences were detected with a recall of 68% in an imbalanced dataset. A CRF model with BERT based features outperformed a fine-tuned BERT model for cause-effect detection with a macro recall of 68%. This led to 96,676 sentences with cause-effect associations. "Diabetes" was identified as the central cluster followed by "Death" and "Insulin". Insulin pricing related causes were frequently associated with "Death". Conclusions: A novel methodology was developed to detect causal sentences and identify both explicit and implicit, single and multi-word cause and corresponding effect as expressed in diabetes-related tweets leveraging BERT-based architectures and visualised as cause-effect-network. Extracting causal associations on real-life, patient reported outcomes in social media data provides a useful complementary source of information in diabetes research.

Perceval Wajsburt, Yoann Taillé, Xavier Tannier

This paper studies the effect of the order of depth of mention on nested named entity recognition (NER) models. NER is an essential task in the extraction of biomedical information, and nested entities are common since medical concepts can assemble to form larger entities. Conventional NER systems only predict disjointed entities. Thus, iterative models for nested NER use multiple predictions to enumerate all entities, imposing a predefined order from largest to smallest or smallest to largest. We design an order-agnostic iterative model and a procedure to choose a custom order during training and prediction. To accommodate for this task, we propose a modification of the Transformer architecture to take into account the entities predicted in the previous steps. We provide a set of experiments to study the model's capabilities and the effects of the order on performance. Finally, we show that the smallest to largest order gives the best results.

Ivan Lerner, Nicolas Paris, Xavier Tannier

We aimed to enhance the performance of a supervised model for clinical named-entity recognition (NER) using medical terminologies. In order to evaluate our system in French, we built a corpus for 5 types of clinical entities. We used a terminology-based system as baseline, built upon UMLS and SNOMED. Then, we evaluated a biGRU-CRF, and an hybrid system using the prediction of the terminology-based system as feature for the biGRU-CRF. In English, we evaluated the NER systems on the i2b2-2009 Medication Challenge for Drug name recognition, which contained 8,573 entities for 268 documents. In French, we built APcNER, a corpus of 147 documents annotated for 5 entities (drug name, sign or symptom, disease or disorder, diagnostic procedure or lab test and therapeutic procedure). We evaluated each NER systems using exact and partial match definition of F-measure for NER. The APcNER contains 4,837 entities which took 28 hours to annotate, the inter-annotator agreement was acceptable for Drug name in exact match (85%) and acceptable for other entity types in non-exact match (>70%). For drug name recognition on both i2b2-2009 and APcNER, the biGRU-CRF performed better that the terminology-based system, with an exact-match F-measure of 91.1% versus 73% and 81.9% versus 75% respectively. Moreover, the hybrid system outperformed the biGRU-CRF, with an exact-match F-measure of 92.2% versus 91.1% (i2b2-2009) and 88.4% versus 81.9% (APcNER). On APcNER corpus, the micro-average F-measure of the hybrid system on the 5 entities was 69.5% in exact match, and 84.1% in non-exact match. APcNER is a French corpus for clinical-NER of five type of entities which covers a large variety of document types. Extending supervised model with terminology allowed for an easy performance gain, especially in low regimes of entities, and established near state of the art results on the i2b2-2009 corpus.

Charlotte Rudnik, Thibault Ehrhart, Olivier Ferret et al.

News agencies produce thousands of multimedia stories describing events happening in the world that are either scheduled such as sports competitions, political summits and elections, or breaking events such as military conflicts, terrorist attacks, natural disasters, etc. When writing up those stories, journalists refer to contextual background and to compare with past similar events. However, searching for precise facts described in stories is hard. In this paper, we propose a general method that leverages the Wikidata knowledge base to produce semantic annotations of news articles. Next, we describe a semantic search engine that supports both keyword based search in news articles and structured data search providing filters for properties belonging to specific event schemas that are automatically inferred.

Xavier Tannier, Nicolas Paris, Hugo Cisneros et al.

Objective: Natural language processing can help minimize human intervention in identifying patients meeting eligibility criteria for clinical trials, but there is still a long way to go to obtain a general and systematic approach that is useful for researchers. We describe two methods taking a step in this direction and present their results obtained during the n2c2 challenge on cohort selection for clinical trials. Materials and Methods: The first method is a weakly supervised method using an unlabeled corpus (MIMIC) to build a silver standard, by producing semi-automatically a small and very precise set of rules to detect some samples of positive and negative patients. This silver standard is then used to train a traditional supervised model. The second method is a terminology-based approach where a medical expert selects the appropriate concepts, and a procedure is defined to search the terms and check the structural or temporal constraints. Results: On the n2c2 dataset containing annotated data about 13 selection criteria on 288 patients, we obtained an overall F1-measure of 0.8969, which is the third best result out of 45 participant teams, with no statistically significant difference with the best-ranked team. Discussion: Both approaches obtained very encouraging results and apply to different types of criteria. The weakly supervised method requires explicit descriptions of positive and negative examples in some reports. The terminology-based method is very efficient when medical concepts carry most of the relevant information. Conclusion: It is unlikely that much more annotated data will be soon available for the task of identifying a wide range of patient phenotypes. One must focus on weakly or non-supervised learning methods using both structured and unstructured data and relying on a comprehensive representation of the patients.

Xavier Tannier, Philippe Muller

Temporal information has been the focus of recent attention in information extraction, leading to some standardization effort, in particular for the task of relating events in a text. This task raises the problem of comparing two annotations of a given text, because relations between events in a story are intrinsically interdependent and cannot be evaluated separately. A proper evaluation measure is also crucial in the context of a machine learning approach to the problem. Finding a common comparison referent at the text level is not obvious, and we argue here in favor of a shift from event-based measures to measures on a unique textual object, a minimal underlying temporal graph, or more formally the transitive reduction of the graph of relations between event boundaries. We support it by an investigation of its properties on synthetic data and on a well-know temporal corpus.