On any given day in America, there are at least one thousand children fighting for their lives in Pediatric Intensive Care Units (PICUs). In the PICU the patient's condition is carefully monitored with automatic sensors. Most of this data is shown in a five-minute "sliding window" display, so a doctor summoned to a patient's bedside always has her most recent history to consider. However what happens to the data that "falls off" this sliding window? In most PICUs, a tiny fraction of it is coarsely aggregated and recorded, but surprisingly, most of this data is simply discarded. Even if most or all the data is recorded, its sheer volume simply overwhelms researchers and analysts; very few tools exist to help them make sense of and learn from this data. This currently discarded data is a potential goldmine of actionable knowledge that could improve outcomes (decreased mortality/morbidity, reduce pain, etc.), and reduce costs (implicit in reduced length of stay). However, the very nature of this data - multivariate, heterogeneous, high dimensional, temporal, noisy, biased, and high frequency - poses significant challenges for traditional analytical techniques from statistics and data mining.
In this project, an interdisciplinary team of investigators is developing: (a) xcalable machine learning algorithms for mining archives of annotated PICU data to find regularities and patterns that can be used to aid in diagnostics and prediction of outcomes; and (b) techniques for monitoring ICU telemetry in real time to detect whether the patterns and rules discovered in the offline step have occurred and can be used to guide interventions (actions by the doctor).
The project brings together experts in data mining (Keogh, Tsotras), high performance computing (Najjar), and medicine (Wetzel) to investigate holistic solutions to the above problems. The project contributes to research-based advanced training of graduate and undergraduate students at the University of California Riverside. The findings, datasets, software, and teaching materials created by this project will be archived in perpetuity at www.cs.ucr.edu/~eamonn/UCRPICU/
