Transcript Hypertension Decision Aid (HDA) - An Internet

My Research Areas
Xudong William Yu
Department of Computer Science
Southern Illinois University Edwardsville
Overview

Artificial Intelligence (most likely to chair)
 Model-based Reasoning and Diagnosis
 Expert Systems
 Integrated Systems
 AI in Education: Robotics, Learning-by-Teaching

Database:


Data Warehouse (could chair)
AI+DB: Data Stream Mining (will chair)
Some Projects





HDA (Hypertension Decision Aid)
MDS (Multi-level Diagnosis System)
MIDST (Mixed Inference Dempster-Shafer
Tool)
DOC (Diagnosis Of Complex systems)
Betty the Brain
HDA Background
Chronic Hypertension

High-blood pressure
 140/90 mmHg or higher
 Approximately 50 million hypertensive in the U.S.
 Number one reason a patient in this country seeks
medical care
 Number one reason physicians in this country
prescribe medication
Background
Primary Hypertension
 “Primary” – no specific cause, contributing factors:
 Genetics
 Diet & Life style
 Personality type
 “Chronic” – requires a life time of treatment
 Treatments include medications, diet changes, life
style changes
Background

Risks of untreated hypertension
 Stroke
 Heart Attack
 Heart Failure
 Kidney Failure
 Cardiovascular Disease

Measuring the effectiveness of treatments depends on
obtaining regular and accurate measurements of a
patient’s blood pressure
Motivation:
Communication is the cornerstone of medical practice
Physicians:
 Listen carefully to patient
concerns
 Take accurate & complete
medical histories
 Help patients understand their
medical problems
 Communicate treatment
recommendations & medical
advice
Patients:
 Relay medical history
 Accurately report signs &
symptoms
 Voice concerns & questions about
conditions & treatments
Poor communication is a major cause of misdiagnosis, poor
compliance of therapy, and malpractice claims.
Communication Characteristics of HDA

Asynchronous – does not require appointments or
time off

Push Medium – does not require a special activity to
find information

Directed Conversation – relevant & personalized
content to the patient’s condition
HDA System Architecture
System Interfaces


Patient Interface
Used to record and
report patients’ vital
signs and stats
Physician Interface
Used to monitor
patient status and
display trend
graphically
Main Components of HDA

Decision Module (DM)
 Use expert system technology to provide decision
support to physicians.

Treatment Database (TD)
 Stores the medical records of patients.

Rule Editing and Verification (REV) Tool
 Provides online assistance to patients.
Decision Module
Main functions:
 Monitor the TD
 Evaluate the treatment plans of patients
 Suggest changes when necessary
Implementation:
 FreeShell Live
two senior projects
Decision Module
Three operating modes:
1. Routine Evaluation
 Periodical evaluation treatment plans
 Basic steps:
 Query the database for the latest data
 Derive a list of medications ranked by Certainty Factor (CF) values
 Takes appropriate action based on the difference in CF value
between current medication and the top-ranked medication:
 Case 1: <0.2, no action
 Case 2: 0.2 – 0.5. Inform the primary care physician.
 Case 3: >0.5, Advice the patient to contact primary care
physician immediately
Decision Module
2. Emergency Evaluation

Triggered by changes in a patient’s vital signs.

Evaluation is similar to routine evaluation.

Inform the primary care physician immediately.
3. Online consultation

Initiated by the primary care physician

Complete evaluation with all test results
Decision Module
Knowledge Base

Based on “The Sixth Report of the Joint National
Committee on Prevention, Detection, Evaluation, and
Treatment of High Blood Pressure”

Use production rules:
Conditions  Conclusion CF, Conclusion CF, …
Ex.,
Total-cholesterol = Above-Normal or
Total-cholesterol = Increasing
Conclude:
Medication = Thiazide -600
Medication = Loop-Riuretics -600
An Example Rule
Knowledge Engineering



Medications are are
ranked by CF values
Certainty Factors
are combined and
fine tuned
A small subset 
CO
SVR
Male
P14
36-45
P14
>65
M5
L19
LDL
L7
HDL
Initial
T11
E
0.5
-0.7
-0.8
0.3
T12
E
0.5
-0.7
-0.8
0.2
T13
E
0.5
-0.7
-0.8
T21
E
0.6
-0.6
-0.8
-0.8
0.3
T22
E
0.1
-0.8
0.2
T23
E
0.1
-0.8
0.2
The REV Tool
Knowledge Acquisition in DM
 A GUI for editing rules
 Provide a vocabulary to reduce syntactic errors.
 Perform Front-end knowledge base verification.
 Test for completeness with cases selected based their
relevance.
REV Interface
Knowledge Base Verification

Local inconsistencies and incompleteness are
detected immediately.

Example of local incompleteness
 Unusable rules:
 Contains conditions that are not verifiable.
 Dead-end rules:
 Not in a path to a conclusion.
Knowledge Base verification
Categories of local inconsistencies:

Redundant rules:



Contradictory rules



A  (C 0.7)
A  B & (C –0.5)
Over-specified conditions:



AB &C
AB &D
AC
A&BC
Circular rules:



AB &C
CD
DA&E
HDA Summary




Supplement the physician & patient relationship with
greater interaction
Help the physician in patient monitoring and decision
making.
Personalizes the content of the interaction similar to
direct visitations
Provides patients an active role in their health care &
immediate feedback (“plug-in” feeling)
Some Related Research

Diagnosis
 Model based vs. Associational
 Modeling for Diagnosis

Database Design
 Extending UML Class Diagram
Model Based Diagnosis
Three Subtasks
 Candidate generation
 Candidate testing
 Candidate discrimination
Diagnosis = Blame Assignment


Conflict: Components involved in the
discrepancy
Example: (M1, M2, A1)
Model Based Diagnosis

Expert system approach uses associational
knowledge


Considered “shallow” & “Brittle”
Model-based diagnosis
Based on a model of the system
 Tend to be more complete
 Key: an adequate model

Pneumatic System
A Method for Candidate Generation
Example Equation Model
v
l
Tho  Thi  Ch (Thi  Tci )( 
)
l Cc Rc D
2
2
R
l
cR
p
p
D  CcChR p2 2 
(CcRc( Rh  Rp)  ChRh( Rc  Rp)) 
( Rc  Rh  Rp)
RhRcl
RcRh
v
l
C c  Vc   tFc c  Fc c
v
A
X  ( Pro  Rs Fs ) E
k
E  E f  Ec (Tset  RcsQr )
X
Rs

A
Fs E
k
Causal analysis on the equations
To relate change in output parameter to changes
in component

Compute partial derivatives on the equations
for X X
A
Rs

k
Fs E

Perform sign analysis, since A, k, E, Fs are +, the
partial derive is -, thus PDC(X, Rs-) = +

Through propagation, we obtain
PDC(Tho , Rs-) = +
Tho Tho Cc Fc X
l
l

 (Ch (Thi  Chi ) 2
)  ( c )  (
Rs
v
Cc Fc X Rs
Cc Rc D
P3
C3
)  (
A
Fs E )
k
Generation Partial Conflicts


Current OBS: Tho above normal
Partial derivative for Tho
Tho Tho Cc Fc X
l
l

 (Ch (Thi  Chi ) 2
)  ( c )  (
Rs
v
Cc Fc X Rs
Cc Rc D

P3
C3
)  (
A
Fs E )
k
Partial conflict for Tho:
PC(Tho+) = Rp+  Ef+  K-  Rs-  Ec  Rcs-
An Integrated Architecture
Knowledge-Base
Editor
User
Interface
Model-Based
Diagnoser
Associational
Module
Knowledge
Base
Inference
Engine
Model
Builder
Diagnostic
Controller
MBD
Module
Some Future Work





Methods for pattern recognition from online,
continuous data stream
Methods for modeling of continuous systems
Create modeling and analysis tools
Knowledge Acquisition: ex., SQL  Rules
GUI Tools