Automatisation de l’analyse des dossiers médicaux pour une prise de décision clinique optimisée

Abstract

This study is situated within the context of the digital transformation of medical records and aims to automate the summarization of clinical notes to support clinical decision-making. The main objective is to develop a model capable of generating coherent and factual summaries from the discharge summaries of the MIMIC-IV dataset. The adopted methodology is based on the fine-tuning of the LongT5 model, chosen for its ability to process very long text sequences. Particular attention was given to data preproces- sing, notably the generation of coherent and precise target summaries to ensure high-quality supervision. The adaptation of the pre-trained model was then efficiently performed using the LoRA (Low-Rank Adaptation) method. Our final model, named MedSum-LongT5, achieves solid performance with scores of 52.6% for ROUGE-1, 35.3% for ROUGE-2, and 42.9% for ROUGE-L. These results signifi- cantly outperform non-specialized baseline models, validating the effectiveness of the approach. Qualitative analysis conf