Quantitative Issues in Cancer Research Working Group Seminar
Kimberly Greco, PhD Student, Department of Biostatistics, Harvard T.H. Chan School of Public Health
Building Graph Neural Networks from the Ground Up: Overcoming Challenges in Disease Prediction from EHR
Abstract: Graph Neural Networks (GNNs) have emerged as powerful tools for disease prediction using Electronic Health Records (EHRs), enabling breakthroughs in identifying latent health conditions and temporal phenotypes. While the core concepts behind GNNs are conceptually straightforward, constructing effective models from scratch requires careful consideration of architecture and hyperparameters. In this talk, we’ll explore the foundational steps of building GNNs, addressing challenges such as overfitting, oversmoothing, and vanishing gradients. Drawing on empirical insights, we’ll discuss strategies for optimizing GNN performance in predictive health applications, bridging the gap between theory and implementation.