LAHVA: Linked Animal Human Health Visual Analytics

1
LAHVA Linked Animal Human Health Visual
Analytics

• Presented by Ross Maciejewski
• Purdue University Regional Visual Analytics Center

2
Visual Analytics

• Visual analytics is the science of analytical
  reasoning facilitated by interactive visual
  interfaces.
• People use visual analytics tools and techniques
  to
• Synthesize information and derive insight from
  massive, dynamic, ambiguous, and often
  conflicting data.
• Detect the expected and discover the unexpected.
• Provide timely, defensible, and understandable
  assessments.
• Communicate assessment effectively for action.

3
Project Goals

• Linked Animal-Human Health Visual Analytics
• Integrated temporal, geospatial, multi-source,
  multi-scale analytic capability
• Systemic biological pandemic, syndromic,
  chem/bio/nuclear surveillance, management, and
  response
• Benefits/ impact
• Earlier detection of some environmental and
  emerging health conditions
• Cross-species monitoring reduces false positive
  rate and could improve spread factor analysis
• By monitoring pets, our system covers areas that
  may not have electronic human health surveillance

4
Outline

• Introduction
• Background
• Motivation
• LAHVA
• Case Studies
• Results
• Conclusions Future Work

5
Introduction
6
Previous Work

• EARS - Early Aberration Reporting System
• ESSENCE - Electronic Surveillance System for the
  Early Notification of Community Based Epidemics
• Biosense

7
Motivation

• Data sources vary widely in accuracy and
  reliability
• Many existing systems do not leverage existing
  messaging and vocabulary standards (e.g. HL7,
  LOINC)
• Current systems generate many false positives
• Many geographic areas lack emergency room
  coverage

8
Motivation (cont.)

• 73 of emerging infectious diseases are zoonotic
  in origin
• Cross system monitoring can provide earlier and
  more reliable detection
• Banfield sees 5 of all pets in the United
  States
• Indiana Network for Patient Care (INPC) consists
  of five major hospital systems in Indianapolis
  and serves more than 390,000 patients per year

9
LAHVA

• Consists of three components
• Data Management
• Statistical Analysis
• Visual Analytics

Visual Analytics Interface
Banfield(Pets)
Relational Database
Public Health Officials
INPC(Hospitals)
Cleaning and Transformation
Statistical Analysis
10
Data Management

• Data Preparation
• Detect and remove errors
• Provide feedback to data providers
• Privacy-Preserving Data Sharing
• Currently uses traditional de-identification
  techniques
• Applying visual abstractions of data

11
Statistical Analysis

• Power Transformation
• Square root versus natural log
• Data Normalization
• Determining an appropriate denominator
• Aberration Detection for Sparse, Dependent Data
• CUSUM (cumulative sum) model
• Bootstrapping
• Quantile Measures
• Seasonal-Trend Decomposition Based on Loess (STL)

12
Power Transformation

• Apply logarithm when mean is proportional to
  standard deviation
• Apply square root when data follows a Poisson
  distribution
• Goal is to make mean independent of standard
  deviation
• Banfield data has many counts of zero, so must
  apply either or
• was found to yield a skid-free
  distribution

13
Data Normalization

• Daily counts vary according to seasonal effects,
  day of the week effects, etc.
• Plots of daily counts are noisy, difficult to
  find aberrations
• For better analysis, we applied a six month
  sliding average window

14
Aberration Detection

• CUSUM
• Used to find deviation from expected value
• Applied in CDC to ER data
• Bootstrapping
• Used as a robust alternative to parametric
  inference
• Idea is to sample data with replacement in order
  to simulate data distribution

15
Aberration Detection

• Quantile Measure
• Used to detect unusual variation of cases in
  retrospective analysis
• For pets within a radius of incidence, we
  identify symptomatic encounters over a time
  window
• There is a distance to the epicenter associated
  with each symptomatic encounter
• An adverse event with a given epicenter should
  cause distribution of these distances to be
  shifted downward

16
STL

• Seasonal-Trend Decomposition Based on Loess
• Time series can be viewed as the sum of multiple
  trend components
• For each data signal, components are extracted
• Can then analyze correlation between components

17
Visual Analytics

• Direct access query to database
• Pre-computed statistical plots
• Factor specification and filtering

18
Visual Analytics
19
(No Transcript)
20
Case Study Effects of Seasonal Influenza (Study
Description)

• Seasonal Influenza is actively studied via
  emergency room chief complaints
• Can equivalent flu-like symptoms in companion
  animals be used as predictors?

21
Case Study Effects of Seasonal Influenza (Study
Results)

• From our results, respiratory illnesses in
  dogs seem to proceed those in humans by 10 days.

22
Case Study Assessing Effects of a Chemical
Release (Study Description)

• In 2006 industrial wastewater was accidentally
  released prompting a public health investigation.
• Results of the investigation were self-reported
  surveys, no emergency room data was available.
• Can companion animal health be used to determine
  the effects of such a release?

23
(No Transcript)
24
(No Transcript)
25
Case Study Assessing Effects of a Chemical
Release (Study Results)
26
Conclusions

• Our system links visual and statistical analysis
  tools to enhance health-care monitoring
• We have provided factor specification and
  filtering components.
• Two cases studies were chosen to assess the
  potential usefulness of such a system and
  data-sources

27
Future Work

• Directly connect the statistical tools we have
  developed into the LAHVA framework.
• Add functionality for users to directly select
  an area for analysis (for example draw a circle
  and determine if the current syndromes in that
  area are exceeding the expected value).
• Add functionality for spatio-temporal cluster
  detection.
• Add functionality for time ghosting of
  individual pet syndromes.

28
Acknowledgements

• Department of Homeland Security
• Banfield the Pet Hospital
• Indiana Network for Patient Care
• National Science Foundation
• David S. Ebert and Tim Collins from Purdue
  University Regional Visualization and Analytics
  Center
• Benjamin Tyner, Cheng Zheng and William
  Cleveland from Purdue University Department of
  Statistics
• Larry Glickman, Nita Glickman, George Moore from
  Purdue University School of Veterinary Medicine
• Rimma Nehme and Mourad Ouzzani from the Purdue
  Cyber Center
• Shaun Grannis from IU School of Medicine