|
Abstract:
|
Consider daily counts of emergency department visits, cross-classified into a list of counts by medical syndrome, hospital, patient age group, zip, and gender. We monitor these many time series for anomalies, such as spikes in visits related to illnesses in specific sub-populations. The Kalman Filter for Contingency Tables, or KFC, provides a framework and specific methodology for modeling and anomaly detection. It includes a loglinear or graphical model for expected counts, which are updated by a Kalman Filter, and shrinkage estimation applied to ratios of observed to expected, adapted from DuMouchel's Gamma Poisson Shrinker for very large tables of counts. KFC is part of ongoing work at AT&T Labs - Research in collaboration with Emergency Medical Associates of New Jersey, and extends new methods for biosurveillance of univariate time series to highly multivariate settings.
|
