Keita Nakao - From Words to Disease Trajectories: Exploring Parkinson’s Disease - ICJ (2/17)

lundi 12 oct. 2026, 10:30 → 11:30 Europe/Paris

Can everyday medical records tell us how a disease develops before it is diagnosed? In this talk, I will introduce an approach to exploring early signs of Parkinson’s disease using electronic health records. These records contain descriptions of symptoms collected over many years, but turning this text into information that we can analyze and interpret is challenging.

To address this, we developed a representation called HCVR that captures two aspects of each clinical entry: which symptom category it belongs to and how favorable or unfavorable its meaning is. This allows us to turn a person’s medical history into a series of numerical profiles while retaining clinically meaningful information. We then use a hidden Markov model to infer underlying health states and explore how people move between them over time.

Using records from approximately 20,000 people with Parkinson’s disease and 19,000 controls, we found that people followed several different pathways before diagnosis. Even among people assigned to the same underlying state, their clinical records contained differences that helped distinguish those who later received a Parkinson’s diagnosis from controls. I will discuss how combining language models with an interpretable statistical model can help us explore the early stages of disease and understand why the path to diagnosis differs between individuals.