The aim of this work is to extract Temporal Entities from patients’ EHR from pediatric hospital specialising in Rare Diseases, thus allowing to create a patient timeline relative to diagnosis . We aim to perform an evaluation of NLP tools and Large Language Models (LLM) to test their application in the field of clinical study where data is limited and sensitive. We present a short annotation guideline for temporal entity identification. We then use the tool EDS-NLP, the Language Model CamemBERT-with-Dates and the LLM Vicuna to extract temporal entities. We perform experiments using three different prompting techniques on the LLM Vicuna to evaluate the model thoroughly. We use a small dataset of 50 EHR describing the evolution of rare diseases in patients to perform our experiments. We show that among the different methods to prompt a LLM, using a decomposed structure of prompting method on the LLM vicuna produces the best results for temporal entity recognition. The LLM learns from examples in the prompt and decomposing one prompt to several prompts allows the model to avoid confusions between the different entity types. Identifying the temporal entities in EHRs helps to build the timeline of a patient and to learn the evolution of a diseases. This is specifically important in the case of rare diseases due to the availability of limited examples. In this paper, we show that this can be made possible with the use of Language Models and LLM in a secure environment, thus preserving the privacy of the patient

This paper presents our two deep learning-based approaches to participate in subtask 1 of the Chemotimelines 2024 Shared task. The first uses a fine-tuning strategy on a relatively small general domain Masked Language Model (MLM) model, with additional normalization steps obtained using a simple Large Language Model (LLM) prompting technique. The second is an LLM-based approach combining advanced automated prompt search with few-shot in-context learning using the DSPy framework.Our results confirm the continued relevance of the smaller MLM fine-tuned model. It also suggests that the automated few-shot LLM approach can perform close to the fine-tuning-based method without extra LLM normalization and be advantageous under scarce data access conditions. We finally hint at the possibility to choose between lower training examples or lower computing resources requirements when considering both methods.