Events

CS Colloquium: Challenges and Advances in Disease Prediction Across Medical Data Modalities

Speaker: Eran Halperin

Location: 60 Fifth Avenue, Room 150
Videoconference link: https://nyu.zoom.us/j/96457147314

Date: Tuesday, March 18, 2025

Predicting disease onset and progression is a fundamental challenge in computational medicine, with important implications for chronic disease
management and treatment planning. While AI methods from other domains have
been applied to disease prediction with some success, several key
challenges make this task uniquely difficult. Unlike many AI/ML problems
where human experts can provide ground-truth labels, disease prediction
often remains unresolved even by clinicians. Additionally, medical data
modalities—such as electronic health records (EHR), medical imaging, and
genomic data—exhibit distinctive characteristics, including systematic
biases, missingness, and interpretability challenges. Without addressing
these factors, predictive models risk being suboptimal, lacking
explainability, or failing to account for critical data limitations.

In this talk, I will present recent advances in disease prediction across
three distinct modalities. Specifically, I will discuss (1) the adaptation
of the GPT architecture for disease prediction using EHR data, (2) the
identification of disease biomarkers from 3D medical imaging, and (3) the
application of methylation risk scores to reduce missingness in EHR data
and improve predictive performance. Through these examples, I will
highlight key methodological considerations and challenges in developing AI
models for medical applications.