Leveraging Generative Large Language Models for Temporal Relation Extraction From French Clinical Narratives:...

Temporal relation extraction (TRE) in clinical narratives is crucial for understanding patient history, disease progression, and treatment pathways. However, it remains challenging due to limited annotated data and to the complexity of clinical text, including domain-specific terminology, inconsistent information, and implicit temporal reasoning. These challenges are particularly acute for rare diseases, where information such as the date of diagnosis or the onset of key phenotypes is rarely available in structured data, despite being essential for estimating diagnostic delay and reconstructing the natural history of the disease.

We explore the potential of open-weight and on-premises large language models (LLMs) on clinical TRE and normalization using zero- and few-shot prompting, aiming to reduce the time-consuming and costly annotation process. Based on recent promising capabilities of LLMs in understanding and reasoning over text, we evaluate whether these models can efficiently perform TRE and normalization in real-world clinical settings.

We cast the temporal relation task as a question-answering task, in which the models extract the temporal expressions associated with a given clinical event. We propose a prompt-chaining strategy that sequentially performs normalization on the extracted temporal expressions. Our LLM-based approach is evaluated on constructed and annotated French real-world clinical narratives, covering temporal relations involving 2 types of clinical events: phenotypes and rare disease diagnoses. Four open-weight LLMs were evaluated across multiple prompt configurations and compared with rule-based and neural baselines using exact-match metrics, an LLM-as-a-judge evaluation, and human validation. We further evaluate our approach by applying it to the English 2012 i2b2 corpus, involving other types of clinical events.

The proposed approach achieves stable extraction performance with only a small number of in-context examples across both clinical event types, reaching a maximum F₁-score of 0.72 for rare disease diagnoses and 0.61 for phenotype events. LLM-as-a-judge evaluation supports these findings by capturing minor temporal variations penalized by exact-match metrics, while human validation further confirms the clinical validity of the extracted relations. On the 2012 i2b2 corpus, the approach remains robust despite increased complexity and minimal supervision. Adding normalization using the prompt-chaining strategy maintains stable overall performance, indicating minimal error propagation and efficient temporal normalization.

Prompting LLMs with a minimal set of in-context examples can achieve strong performance for TRE and normalization in low-resource settings with limited domain-specific annotations. Relying on open-weight, on-premises models, the proposed approach therefore offers a practical path toward deployment in real-world clinical workflows and could further be extended to extract key clinical information, such as age at diagnosis or age at onset, particularly in the context of rare diseases, where it could help prefill structured reporting forms, reducing the manual effort needed to retrieve such information directly from clinical narratives.