Soutenance de thèse Adam Remaki

Titre: Structurer les notes cliniques pour la réutilisation des données de santé : extraction et normalisation de l’information médicale par traitement automatique des langues

Title: From Clinical Notes to Structured Data for Secondary Use: Extracting and Normalizing Medical Information with Natural Language Processing

Lieu: Limics, campus des Cordeliers

Abstract:

This thesis investigates how Natural Language Processing (NLP) can transform unstructured clinical text into structured variables to support the secondary use of Electronic Health Records (EHRs). EHRs contain large amounts of data, but only part of this information is available in structured and directly queryable tables. Yet, a large part of clinical information remains in clinical notes, which provide richer descriptions of patients but must first be structured before they can be reused at scale.

The thesis addresses this challenge through three complementary contributions. First, it studies an end-to-end NLP pipeline designed to structure information from discharge summaries in the Greater Paris University Hospitals clinical data warehouse. The pipeline extracts laboratory tests and drug treatments, interprets their clinical context, including negation, uncertainty, conditional statements, and temporality, and links them to standard concepts so that they can be combined with already structured EHR data. This work was evaluated in a cohort of 6,891 patients with immune-mediated inflammatory diseases and applied to 22,194 discharge summaries. The pipeline reached an end-to-end F1-score of 71.1 for laboratory test normalization and 89.3 for drug name normalization. At the cohort level, adding discharge summaries identified positive antinuclear antibodies in 53.6% additional patients with systemic lupus erythematosus compared with structured data alone.

This first contribution shows the practical value of structuring clinical text, but also highlights a key limitation: extracted mentions must be reliably linked to standard concepts. The thesis therefore focuses next on biomedical entity linking (BEL), which maps textual mentions to concepts in a knowledge base (KB). In the first pipeline, BEL mainly relied on the mention itself, which can fail when the same surface form is ambiguous and the correct concept depends on context. Context-aware BEL models can address this problem, but they require costly expert annotations that cover only a small fraction of the target KB. To reduce this bottleneck, the thesis introduces SynCABEL, a synthetic contextualized augmentation framework. SynCABEL uses large language models to generate contextual examples from KB concepts and combines them with human annotations to train generative BEL models. Experiments on multilingual benchmarks show that SynCABEL improves BEL, reaching an average Recall@1 of 75.4 with Llama-3-8B and guided inference, and matching full human supervision with up to 60% fewer annotations.

Third, the thesis studies document-level consistency in BEL. Most approaches link each mention independently, often using only the local sentence, even though biomedical documents frequently contain repeated concepts, abbreviations, and complementary cues. To address this limitation, the thesis introduces LongBEL, a document-level generative framework that conditions each prediction on the full document and on a memory of previous linking decisions. This allows the model to reuse information introduced earlier in the document when linking later mentions. Across three datasets in English, French, and Spanish, LongBEL outperformed the sentence-level model. The gains were largest for concepts already mentioned in the document and when such recurrences were sufficiently represented in the training data.

Overall, this thesis shows that NLP can transform part of the information contained in free-text clinical notes into structured variables that complement existing EHR data for secondary use. It also addresses two key limitations of BEL: annotation scarcity and the lack of consistency across predictions within the same document. Finally, the reuse of the end-to-end pipeline across at least seven clinical research projects illustrates its practical value beyond the methodological evaluations presented in this thesis.