Spatiotemporal Cluster Detection in ESSENCE Biosurveillance Systems

1
Spatiotemporal Cluster Detection in ESSENCE
Biosurveillance Systems

• Panelist Howard Burkom
• National Security Technology Department,
• John Hopkins University Applied Physics
  Laboratory
• DIMACS Working Group
• Workshop on Analytical Methods for
• Surveillance of Multidimensional Data
  StreamsRutgers University, Piscataway NJ
• February 19, 2004

2
Problem/Data Context of ESSENCE Surveillance
Systems
Absentee Rates
Sales of OTC Remedies
Hosp ER Admissions
Physician Office Visits
Normalization Analysis Fusion
Counts/Clusters of Statistical Significance
Who? What? Where? When?
Epidemiological Significance
3
Applying Statistical Process Control to Multiple
Data Streams

• Multiplicity from intertwined effects multiple
  data sources, regions, strata (syndrome groups,
  product groups)
• Multiple univariate methods
• Critical issue use individual detector outputs
  without getting overwhelmed by multiple testing
• Low power for anomalies spread over inputs
• Multivariate methods
• Critical issue need modifications to reduce
  alerts due to irrelevant changes in data
  relationships
• Need to retain power in individual source data

4
Significance Assessment Multiple Univariate
Alerting Algorithms

• Bonferroni bound replace a by a/N
• Alert based on individual outputs (conservative)
• Edgingtons consensus method (1972)
• Combined prob from alg. comb .of N individual
  p-values
• Z-score approximation
• ( mean(p-values) 0.5 ) / ( 0.2887 / vN )
• Bayes Belief Net
• Originated effort to add sensor data,
  intelligence info,
• Recently applied to separate algorithm outputs
• Can weight each type of information based on
  training data and/or intuition
• Configurable to soften thresholds for evidence
  accrual

5
Multivariate Alerting Strategies

• Variants of Hotellings T2
• m vector mean est. from current baseline
• S est. of covariance matrix calc. from baseline
  
• X multivariate (filtered?) data from test
  interval
• T2 statistic (X- m) S-1(X- m) (Ye et al, 2002)
• Neighbor-regression preconditioning strategy of
  Hawkins removal of covariance effects
• MEWMA (Lowry), MCUSUM (Crosier,
  Pignatiello/Runger)
• Numerous strategies, adaptations to Poisson data
• But which is appropriate for multivariate
  syndromic data streams?
• Can EWMA/Shewhart (or CUSUM/Shewhart) encompass
  both point-source bioweapon epicurve and
  seasonal endemicgtepidemic outbreak?

6
Detection Challenge faint rise in all 3 data
sets
Respiratory Syndrome Data Counts
Military Dx
Military Rx
Civilian Dx
7
Detection Challenge faint rise in all 3 data
sets
Respiratory Syndrome Data Counts
Lowrys MEWMA Day 4 alert at each FA rate
8
Scan Statistics for Biosurveillance
Scarlet Fever Outbreak Study
Analysis of Claims Data in National Capital Area
ICD9 codes for scarlet fever 034 034.1
10 cases, 5 days p 0.013
15 cases, 12 days p 0.002
11 cases, 7 days p lt 0.001
9
Surveillance combining outpatient visits, OTC
anti-flu sales, school absenteeism
10
Practical Issues in Spatiotemporal Monitoring and
Evaluation

• Control needed for mismatched scales variances
  among data sources
• To retain power in indiv. sources, gain combined
  sensitivity
• Difficult to assess delays, relative scales of
  effects among separate sources, in both
  background signal
• Simulation much harder to validate
• If distance matrix is used, it should reflect
  proximity according to the epidemiological case
  definition
• Modifications to reflect plausible demographic
  behaviors
• The importance of significance testing grows
  with the number of sources, especially for
  subregions where expected counts are low
• More sources gt more small spurious clusters

11
Finding Clusters with Multiple Data Sources

• For candidate cluster J1, the Kulldorff
  likelihood ratio is
• LR(J1) (O1/E1)O1 ((N-O1) / (N-E1)) (N-O1)
• where O1 number of cases inside J1,
• E1 number of cases outside J1,
• N total case count
• Extension by treating multiple sources as
  covariates
• O1 SO1k, E1 SE1k, N SNk, for sources
  k1,,K
• adjusted method problem of adding sources
  with mismatched scales, variances
• Alternate multisource approach stratified
  scan statistic
• S log( LR(J1k) ), k1,,K
• reduces chances for a noisy source to overwhelm
  others
• can cost power to detect faint signal spread
  over sources

12
FROC Performance Assessment Adjusted vs
Stratified Multisource Scan Statistics
Prob. Signal-Based Significant Cluster
Prob. Random Background Significant Cluster