Project Mimic: Simulation for Syndromic Surveillance Thomas Lotze Applied Mathematics and Scientific Computation University of Maryland Galit Shmueli and Inbal Yahav RH Smith School of Business University of Maryland with Howard Burkom and Sean Murphy JHU Applied Physics Lab This work was partially supported by NIH grant RFA-PH
Outline The Biosurveillance Problem Motivation: Reasons for simulation Simulation Methodology Options/Generation Mimicking a dataset Analysis Is this is a good mimic? Results
The Biosurveillance Problem
The Biosurveillance Problem, cont. Given time series (usually pre-diagnostic daily data) Detect disease outbreaks With few false alerts Early
Difficulties with Biosurveillance Data Teams work on different authentic datasets Each team has their own private data Cannot compare results Researchers with no data cannot join the effort Data are unlabeled We don’t know exactly when there are outbreaks Challenges evaluation of algorithm performance Hinders comparison of different algorithms
Project Mimic Q: What if there was a way to generate pseudo-authentic data similar in statistical structure to real data AND insert simulated outbreak signatures into it? A: we’d have new, labeled pseudo-real data!
Project Mimic: Dataset Mimicker “Mimics” statistical structure of background data Levels of counts of different series Day-of-week patterns Seasonal patterns Holidays Within-series autocorrelation Cross-series cross-correlation Extracts features from the authentic dataset Output: dataset that “looks” like real dataset
Set of 6 series from one city OriginalMimicked Resp GI
3 series from one city, zoomed in
Mimic Methodology Our method(s): Create random autocorrelated multivariate data Normal or poisson Uses mean, standard deviation, reduced cross- correlation, 1-day acf from original Holiday factor Seasonal factor Day-of-week factor Details at Mimicking implicitly uses a generative model What is the right model?
Evaluating Mimics Test: could the original data have been generated from the mimicker? Compare different generative models If the model were simple, could use AIC Instead, Chi-squared
Chi-squared Goodness-of-fit Tests By series By day of week Separate values into bins Chi-squared Test on counts
Example of Disparity
Project Mimic: Outbreak signature simulator Generates multivariate outbreak-signatures Options: Number of outbreak-signatures in series? Magnitude of outbreak? How many (and which) series will include outbreak- signatures? Stochastic/fixed? Include effects such as DOW, holidays, etc.? (like background data) Output: matrix of outbreak-signatures to be inserted in the background data
Outbreak labels
Project Mimic Combining the background matrix + outbreak-signature matrix yields labeled data Two final products Mimicker: Data and outbreak-signature simulators (in freeware R) Can be used by data owners to disseminate pseudo-data Can be used by research teams to evaluate robustness of methods Mimics: Datasets that mimic DARPA BioALIRT data Benchmark datasets for comparison across groups Can be used to perform optimization methods for improved detection Available at Example: BioALIRT data on 3 series (Resp from civilian/military/prescriptions)
Mimicked data + outbreak-signature
Conclusions Mimic opens the door to: new techniques new researchers First data sets of their kind Open methodology Publicly available Realistic