Extracting Temporal Relations from Patients' EHR with LLMs
By Judith Jeyafreeda

Judith Jeyafreeda will give a talk on extracting temporal relations from patients’ EHR with LLMs

Abstract

The aim of Judith Jeyafreeda’s research is to extract temporal relations from patients’ electronic health records (EHR) at a pediatric hospital specializing in rare diseases, allowing the creation of a patient timeline relative to diagnosis. Her work involves a combination of natural language processing (NLP) tools and large language models (LLMs) fine-tuned for clinical studies, where data is both limited and sensitive. She will present a short annotation guideline developed for temporal relation extraction and a pipeline created for clinical temporal extraction. For this purpose, she uses GLiNER, a BERT-based named entity recognition (NER) tool for extracting temporal entities, and OpenNRE, a BERT-based relation extraction tool.

In terms of using LLMs, Judith has conducted experiments using three different prompting techniques on the LLM Vicuna to thoroughly evaluate its effectiveness in extracting temporal entities. A small dataset of 50 EHRs, describing the evolution of rare diseases in patients, was used for the experiments. The results show that, among the various methods to prompt an LLM, using a decomposed structure of prompting with Vicuna yields the best results for temporal entity recognition. She also applies prompt-based methods for relation extraction, using Llama 3. Moving forward, she will continue to integrate GLiNER, OpenNRE, and LLMs to create a more efficient pipeline for temporal relation extraction.

Biography

Judith Jeyafreeda is a postdoctoral researcher at the University of Paris Cité. She is involved in research around Clinical NLP with the Institute Imagine. She has worked in the same area of research at the University of Manchester, where she developed methods for de-identification of clinical letters and named entity recognition (NER) of diagnoses, followed by linking diagnoses to SNOMED-CT codes. She completed her PhD at the University of Caen, Normandy, where her research focused on “Task-oriented web page segmentation,” developing methods for skimming and scanning web pages for the visually impaired. Her research is centered on various NLP applications.