Artificial Intelligence and the Two Languages of Medicine

Andrew Loza, MD PhD Instructor in the Departments of Biomedical Informatics and Data Science and Pediatrics and Yale University

(Speaker will be presenting via zoom)

Join by Zoom

Meeting ID: 991 2985 2010         Passcode: 923472

Real-World Data holds significant promise to advance personalized medicine and improve our understanding of disease development and treatment response. Despite this promise current analytical methods fall short in key capacities including the ability to integrate the breadth of clinical data which defines a patient’s health state, and the requirement to create a new model for each question. Medical foundation models are deep learning models of healthcare data that are inspired by recent advances in large language models and offer a step forward in addressing these limitations. Instead of operating on sequences of words, medical foundation models operate on the longitudinal sequence of observations, actions, and the passage of time recorded in electronic medical records. Once trained, a medical foundation model can be used to predict any outcome present in the training data without task-specific fine-tuning. During this talk we will take a historical review of the methods and insights which led to this approach. We will then discuss Curiosity, the largest medical foundation model to date trained on 118 million patient timelines. Finally, we will examine current research to improve the performance of these models.