Process design for interpreting deep learning model decisions using large language models

EBSCO and McMaster University’s Health Information Research Unit (HiRU) collaborate to enhance the delivery of high-quality, clinically relevant content to healthcare professionals. Through DynaMed, EBSCO’s evidence-based clinical resource, and HiRU’s literature surveillance program, McMaster PLUS, they provide a continuous stream of rigorously appraised articles from PubMed to clinicians globally and to DynaMed expert authors. Machine learning models have been developed by HiRU to expedite access to this quality evidence, but we do not understand how decisions are made by the models when they classify evidence. To adhere to transparency and responsible AI practices, this project will test whether large language models, like GPT-4o, can provide explanations for the machine learning model decisions. We will test GPT-4o with a series of prompts and approaches and evaluate how well it performs the task. The goal is to develop a process to be implemented within the support services for DynaMed authors to increase their trust in the model decisions.

Faculty Supervisor:

Cynthia Lokker

Student:

Partner:

EBSCO Health

Discipline:

Computer science

Sector:

Information and cultural industries; Retail trade

University:

McMaster University

Program: