Talk of Jesus Lovon

Electronic Health Record (EHR) tables pose unique challenges, among which is the presence of hidden contextual dependencies between medical features with a high level of data dimensionality and sparsity. We present an extensive evaluation of the abilities of LLMs to comprehend EHRs for patient data extraction and retrieval. We conduct extensive experiments using the MIMICSQL dataset to explore the impact of the prompt structure, instruction, context, and demonstration of two backbone LLMs, Llama and Meditron, based on task performance. We propose guidelines that we believe would help the design of LLM-based models to support health search. In a second work, we introduce a new end-to-end training approach for LLM for clinical prediction based on EHR rewriting.

Speaker :
Jesus Lovon is currently a Postdoc Researcher at the Groupe de Recherche en Informatique, Image et Instrumentation de Caen (GREYC), where he works on the application of generative AI to support medical practitioners as part of the PARTAGES project. Before, he was a Postdoc Researcher at the Institut de Recherche en Informatique de Toulouse (IRIT) working on the early detection of major congenital malformation risk in pregnancy, in the context of the INuTERO project. He did his PhD at IRIT - Université Paul Sabatier under the supervision of Lynda Tamine and José G. Moreno. His PhD research topic focused on the integration of structural knowledge into Pre-trained Language Models. Before his PhD, he earned his Master’s degree at Université Paul Sabatier and a Bachelor’s degree from the Universidad Nacional de Ingeniería in Lima, Perú. He have had the chance to do some different research internships at IRIT (Toulouse), and SmartCityPeru Lab (Lima). Also, he worked as a software developer at Wolfram Research in Lima. His research interests are centered around the fields of Natural Language Processing, Information Retrieval, Computation Linguistics, and Healthcare applications