When personalization harms performance

Clinical prediction models often include explicit group attributes such as sex, age and HIV status for personalization (i.e., to target heterogeneous subgroups). In this talk, I will describe how common approaches to personalization can induce "worsenalization" at a group level — i.e., by targeting groups in unnecessarily inaccurate ways. I will discuss how these effects violate our basic expectations of personalization and present work to address them through model evaluation and development. 

Speaker: Berk Ustun, PhD, Assistant Professor, Department of Computer Science and Engineering, UC San Diego

ZOOM Info: Contact Liz Buggs to request Zoom link for seminar