Real-world Patient Data Science

Our research focuses on developing machine learning methods for real-world healthcare data, with particular attention to data that are longitudinal, heterogeneous, incomplete, or imperfectly labeled. I am interested in how these challenges affect risk prediction and decision support in healthcare and public health.

One major area of my research is dynamic clinical risk prediction. Using longitudinal Medicare claims data, I study how a patient’s long-term medical and prescription history can be incorporated into machine learning models to estimate short-term opioid overdose risk at individual prescription events. I am also investigating positive-unlabeled learning and related approaches for healthcare problems in which observed clinical labels may be incomplete and the absence of a recorded diagnosis does not necessarily represent a true negative.

Another area of my work focuses on AI-assisted healthcare and community decision support. I am involved in developing methods to organize and integrate information from community stakeholder interviews into shared pathway maps that can support the identification of service gaps and planning priorities.

Overall, my goal is to develop reliable and useful AI methods that can translate imperfect real-world health data into actionable information for healthcare decision-making.