Hospital intensive care units (ICUs) are particularly interesting medical microcosms as their data are relatively frequent and some types of outcomes are more quickly known. They provide a glimpse into the future challenges and opportunities that await as health data collection becomes more pervasive.
In this talk, I will discuss our recent work in modeling pediatric ICU data using a marked point process model, the piecewise-constant conditional intensity model (PCIM). We apply PCIMs and co-clustering to find data-driven patient clusters from 10,000 patient episodes over 10 years at Children's Hospital Los Angeles. Our clusters are better enriched with respect to mortality than other clustering methods and methods using discrete-time models.
To finish the talk, I will present our new general inference method. Built on Poisson processing thinning, the reversible-jump auxiliary Gibbs sampler allows general posterior distribution inference over a wide class of missing data patterns. I will conclude with some results on video activity understanding.