MAITA Monitoring, Analysis, and Interpretation Tool Arsenal

1
MAITAMonitoring, Analysis, and Interpretation
Tool Arsenal
Jon Doyle William Long Peter Szolovits Philip
Greenspun Christine Tsien
Isaac Kohane Cungen Cao
Clinical Decision Making Group Laboratory for
Computer Science Massachusetts Institute of
Technology
Harvard University Medical School and Childrens
Hospital (Boston)
http//www.medg.lcs.mit.edu/projects/maita
2
Outline

• Participation
• Both integration teams on movement analysis
• Knowledge acquisition from SMEs
• Discussions on battlespace ontology,
  problem-solving methods, evaluation methods,
  OKBC, CP refinements
• Progress
• Monitoring Architecture
• Movement Analysis
• Plans
• Refinement
• Integration
• Decision Models

3
System Overview

• Distributed Monitoring Environment
• Multiple developers and operational users
• Distributed editing tools and libraries
• Multiple knowledge bases and ontologies
• Multiple signal and data sources
• Dynamic interacting monitor processes

• Libraries of problem-solving methods
• Signal transducers and trend detectors
• Control, alerting, and display models

User added monitors and knowledge
Rapid distributed construction
4
Distributed Monitoring Architecture
Control Panels Monitors of Monitors (MoMs) Java
applets
Displays Strip charts Maps Java applets
Network of Monitor Processes wired together by
terminals
HTTP
HTTP
Sockets
Control Terminal
Sockets HTTP OKBC ODBC
Monitor Process
Packets (structured reports)
Input Terminals
Output Terminals
Internal State
Subprocesses and terminal connections
DBs MoM tables Monitor tables
KBs Correlators Transducers Monitor
Library Packet Library Alerting Models Monitor
Knowledge
ODBC
OKBC
Editors Network Structure, Monitor Descriptions
5
Movement Analysis

• Geometric algorithms for road-finding, etc.
• Java and Lisp implementations used by several
  groups
• Suite of elementary MTI signal transducers
• Vehicle histories and classifications
• Starts, stops, enter road, leave road, military,
  nonmilitary
• Convoy identification, starts, stops, tracking
• Classification for certain displacement convoys
• Off-road site detection and population monitoring
• Radar site identification
• Strategy
• Online operation and report generation to intel
  officer
• Report partial information first, refine later

6
Movement Analysis Processes
Traffic Trends
Convoy Motion Display
MTI Tracks
Convoy Motions
Vehicle Motions
Convoy ID
Displaced Sites
Off-road Sites
Displacement Convoy ID
ISE Reports
Radar Sites
Military Vehicles
7
Successes and Failures

• Successes
• Analysis at 10-15X real time
• Good rejection of civilian traffic
• Good identification of convoys
• Good identification of sites
• Reasonable differentiation of military vehicles
• Plausible identification of certain displacement
  motions

• Failures
• Possible errors in understanding SME parameters
• Undiagnosed monitoring failures
• Disappointments
• Task required extensive low-level 2½D
  signal-processing mechanisms involving little
  explicit knowledge and little overlap with
  familiar 1D signals

8
Lessons Learned

• Efficacy of key facts
• Off-road movements
• Coordinated movement histories
• Importance of evaluation semantics
• Common terms in KB language and organization
• Reporting at increasing levels of detail
• Importance of integration
• Integrate early and often
• Consensus architecture
• Simulation and evaluation tools from day 1
• System administration burden and feasibility of
  integrator administration
• Combining inputs as an HPKB component task

9
Plans

• Refine architecture
• Integrate with KB editing environments
• Investigate use in multiple domains
• HPKB domains (battlespace, information assurance)
• Medical (ICU, diabetes, heart disease)
• Formalize decision-making knowledge for HPKB
  alerting tasks

10
Decision Models for Alerting

• Who, when, and how to notify
• Aggregating or serializing related alerts
• Focusing on most important alerts
• Preference representation and reasoning
• Model qualitative and quantitative preferences
  and tradeoffs
• Compile qualitative and quantitative information
  into quantitative utility measures
• KB taxonomies for organizing alerting models and
  their associations with monitoring methods