Progress Toward Earlier Detection: Alerting Algorithms

1
Estimation of Late Reporting Corrections for
Health Indicator Surveillance

Howard Burkom1, PhD Yevgeniy Elbert2, MSc LTC
Julie Pavlin2, MD MPH Christina Polyak2,
MPH 1The Johns Hopkins University Applied
Physics Laboratory 2 Walter Reed Army Institute
for Research Global
Emerging Infections Surveillance Response
System San Francisco, CA November 17,
2003 American Public Health Assoc. 131st Annual
Meeting
2
ESSENCE An Electronic Surveillance System for
the Early Notification of Community-based
Epidemics

• Earlier detection of aberrant clinical patterns
  at the community level to jump-start response
• Rapid epidemiology-based targeting of limited
  response assets (e.g., personnel and drugs)
• Communication to reduce the spread of panic and
  civil unrest

3
ESSENCE Biosurveillance Systems

• Monitoring health care data from 800 mil.
  treatment facilities since Sept. 2001
• System receives 100,000 patient encounters per
  day
• Adding, evaluating new sources
• Civilian physician visits
• OTC pharmacy sales
• Prescription data
• Expanding to nurse hotline, absenteeism data,
  animal health,
• Developing implementing alerting algorithms

4
Using Lagged Data Counts for Biosurveillance

• ESSENCE II data gt hypothesis that earlier stages
  of an outbreak may be more detectable in office
  visit (OV) data than in emergency department data
  
• Depends on existence, duration of typical
  prodrome for underlying disease
• How to exploit this for earlier alerting?
• BUT, our electronic OV data is reported variably
  late, depending on individual providers
• QUESTION How can a timely source of data with a
  reporting lag be used for biosurveillance?

5
Reporting of Civilian Office Visits
Daily Regional Civilian Diagnosis Counts
Respiratory Syndrome Group
6
Office Visit Reporting Promptness by Data Source
Use of Kaplan-Meier Failure Function Curves to
Represent Reporting Promptness
7
Using Lagged Data for Biosurveillance

• Approaches
• Two steps estimate actual counts, apply
  algorithm
• use recent promptness functions by day-of-week,
  other covariates
• apply lateness factors to recent counts
• Brookmeyer R, Gail MH, AIDS Epidemiology A
  Quantitative Approach. New York Oxford
  University Press 1994 Chapter 7
• Use historically early reporting providers as
  sentinels
• Combined approach use regression on counts with
  date and lag as predictors to determine whether
  recent reported data are anomalous
• Zeger, SL, See, L-C, Diggle, PJ, Statistical
  Methods for Monitoring the AIDS Epidemic,
  Statistics in Medicine 8 (1999)
• Linear regression using number of providers
  reporting each day

8
Reporting of ER/Outpatient Visits
Outpatient 80 reported by day 3
ER 50 reported by day 3
Apparent difference in reporting promptness
between ER and other clinics
9
Reporting of Civilian Office Visits21-day
adjustment Week 1
10
Using Provider Counts to Adjust for Lagged
Reporting

• Concept (applied in recent DARPA eval.)
• tabulate doctors or clinics reporting each day
• use residuals of linear regression of daily data
  counts on providers
• accounts for known unknown dropoffs by
  computing actual counts vs expected, given daily
  providers
• can include additional predictor variables
• Can apply process control alerting algorithms to
  residuals
• Significantly attenuates day-of-week effect

11
Counts of Clinic/MTF PairsMilitary Outpatient
Visit Data
City-Wide Respiratory Diagnosis Counts
Number of Clinics Reporting Explains away
unexpected data dropoffs
12
Effect of Provider Count Regression
Visit Counts and Residuals Day-of-Week Effect
Attenuation
Rise due to outbreak?
13
Effectiveness in DARPA Outbreak Evaluation
Challenge