CS Colloquium

Time and Location:

Speaker:

Link:

Machine learning systems in healthcare often fall short of their promises. I argue that a central reason is our inability to precisely characterize how humans and algorithms behave in the real world. Without models to characterize such behavior, our ability to design machine learning systems that effectively complement human judgment is limited.

In this talk, I will present two methods to reason about human and algorithmic decision-making in healthcare.  The first studies how human behavior shapes healthcare data, focusing on underdiagnosis – a pervasive source of corrupted labels in which diseases remain undocumented. I introduce a method to quantify underdiagnosis, revealing systematic gaps in recorded diagnoses and enabling new ways to study how social factors distort observed health data. The second addresses the challenge of understanding algorithmic behavior with limited labeled data, a common challenge in healthcare settings. I show how we can take advantage of unlabeled data to estimate real-world performance efficiently, expanding the contexts in which we can reliably evaluate algorithms. Together, these works illustrate a simple principle: through new methods to carefully characterize the data and algorithms at hand, we can build machine learning systems capable of improving care.