BIDS Monthly Seminar

Modeling time series of data that are stochastic and irregularly sampled remains a challenge across multiple fields. This is especially true in medicine, where numeric and categorical observations are made in a sparse manner across numerous classes at irregular intervals. Time series models well suited for categorical inputs are often not ideal for numeric inputs, and vice versa. Furthermore, sampling of systems is often not performed at random, and measurement time may reflect important information about the evolution of the underlying system. Generative modeling has led to state-of-the-art performance in language processing, image analysis, speech recognition, and more. This talk describes work adapting transformer-based architectures to model Electronic Heath Record data. Two models will be presented, one focusing on adapting a discrete token framework to include both numeric and continuous inputs, and a second focusing on incorporating sequence level properties into the pretraining process for effect estimation and guided decoding.

Andrew Loza received his PhD from Washington University in St. Louis in biophysics studying mechanisms of collective cell migration using time lapse microscopy coupled with computer vision methods and simulation. He completed his MD degree at the Yale University School of Medicine and residency in Internal Medicine – Pediatrics also at Yale. He then completed a Clinical Informatics fellowship in the ACGME Yale/VA program with clinical work at the Yale Internal Medicine – Pediatrics Clinic. Dr. Loza is a physician-scientist whose research focuses on combining classical statistical methods and deep learning to improve disease diagnosis and the delivery of care. His current research focuses on developing statistical and deep learning models of electronic health record data to improve risk prediction, diagnosis, and disease phenotyping.