September 28, 2026 
Noon – 1 p.m. via Zoom
REGISTER NOW »

Session Description

Electronic health record data offer valuable opportunities for prediction in clinical and population health research, but selecting a useful model involves more than maximizing predictive accuracy. This presentation will discuss the balance between predictive performance, interpretability, and practical application when developing models using EHR data. Particular attention will be given to the Feasible Solutions Algorithm, an interpretable variable-selection approach that identifies parsimonious predictive models, and how its results can be compared with more flexible machine-learning methods such as Random Forest. These concepts will be demonstrated through an application to colorectal cancer screening prediction in an Appalachian healthcare population. The presentation will highlight how the intended use of a model can guide decisions about model complexity, performance metrics, and interpretability in clinical and translational research.

Presenter

Maliha Mehnaz Mitu, MS
PhD Candidate, Department of Biostatistics
University of Kentucky College of Public Health