Challenges in Machine Learning for Clinical Decision Making

Abstract

Demo Lecture:

For decades researchers have sought to use machine learning in clinical decision making: aiding or standing in for doctors in making diagnoses and determining treatment. Today, while machine learning is being used to improve healthcare in everything from patient recruitment for clinical trials to robot-assisted surgery, the widespread use of machine learning for clinical decision making has not been realized. Models are often able to make accurate predictions on existing medical records, but barriers such as concept drift, biased data, unexpected feedback loops, and legal challenges have caused models to have negligible or even negative effects on patient outcomes when implemented. Furthermore, even if average patient outcomes improve, that improvement may not reach patients who are demographically dissimilar to the majority.

In this lesson, after defining machine learning we will learn how machine learning models are evaluated, and how some performance metrics can be misleading if used in the wrong situation. We will see how this applies to machine learning for clinical decision making and learn about specific challenges in medical data and clinical practice make model training and implementation difficult. Finally, we will see how machine learning models, if not implemented carefully, can magnify existing inequities in healthcare and metrics data scientists have used to attempt to quantify inequities in machine learning performance.

Bio:

Dr. Chris Magnano earned his bachelor’s in Computer Science and Biology from Swarthmore College in 2014, and his master’s in Computer Sciences from the University of Wisconsin-Madison in 2016. After working as a software developer for Epic Systems, he returned to UW-Madison and completed his PhD in Computer Sciences in 2022. He performed his graduate work in Anthony Gitter’s lab at the Morgridge Institute. Dr. Magnano’s graduate work focused on making computational analyses more accessible to biological researchers, both through creating new education resources and through streamlining biological network analyses. His work in science education ranges from being Environmental Director at YMCA Camp Belknap to university teaching and educational research. He joined the Center for Computational Biomedicine at Harvard Medical School as a curriculum fellow in March of 2022.