Large amounts of multivariate time series data are now being collected for tracking phenomena that evolve over time. Approaches that can incorporate expert biases to discover informative representations of such data are valuable in enabling exploratory analysis and feature construction.
In this talk, we will discuss priors to cluster time series that leverage two different sets of assumptions frequently present in real world data. First, we want to identify repeating “shapes”. For example, when we walk vs. kick, our joint angles produce different repeating shapes over time. How do we discover such repeating shapes in the absence of any labeled data? The second assumption we will tackle is when the generated data does not come from a fixed set of one or two classes as is the case in binary prediction. For example, in clinical data, no two patients are alike. Assuming that generated data is sampled from an m-class distribution can inappropriately bias your learned model.
I will show results from multiple domains but will focus on a novel application of modeling physiologic data from monitoring infants in the ICU. Insights gained from our exploratory analysis led to a novel risk prediction score that combines patterns from continuous physiological signals to predict which infants are at risk for developing major complications. This work was published on the cover of Science Translational Medicine (Science’s new journal aimed at translational medicine work), and was covered by numerous national and international press sources.
Speaker Biography
Suchi Saria recently joined Johns Hopkins as an Assistant Professor in the departments of Computer Science and Health Policy & Management. She received her PhD in Computer Science from Stanford University working with Prof. Daphne Koller. She has won various awards including, a Best Student Paper awards, the Rambus Fellowship, the Microsoft full scholarship and the National Science Foundation Computing Innovation Fellowship. Prof. Saria’s interests are in graphical models, machine learning with applications in modeling complex temporal systems. In particular, she wants to help solve the trillion dollar question of how to fix our health care system and show how these approaches can help us improve the delivery of healthcare.