Transcript Computer Support for Improving Accessibility of Breast

Methods for
Computer-Aided Design and
Execution of Clinical Protocols
Mark A. Musen, M.D., Ph.D.
Stanford Medical Informatics
Stanford University
Research problems
in medical informatics involve
Formulation of models of clinical tasks
and application areas
 Representation of those models in
machine-understandable form
 Development of new algorithms that
process domain models
 Implementation of computer programs
that use models to automate clinically
important tasks

Protocol-based care is
everywhere
Algorithms for mid-level practitioners
 Clinical-trial protocols
 Clinical alerts and reminders
 Clinical practice guidelines

Some basic beliefs
Computer-based patient records
eventually will become ubiquitous
 Clinical protocols can—and should—be
authored from the beginning as
machine-interpretable documents
 Electronic protocol knowledge bases
will allow computer-based patient
records to enhance all components of
patient care and clinical research

Work in protocol-based care

ONCOCIN (1979–1988)
 Clinical

Therapy Helper (1989–1995)
 Clinical

trials in oncology
trials for HIV infection
EON (1989–)
 Reusable
components for automation of
protocols and guidelines in a variety of
domains
Our research addresses

Development of computational models
of
 Planning
medical therapy
 Determining when therapy is applicable
 Reasoning about time-ordered data

New approaches for acquisition,
representation, and use of medical
knowledge within computers
EON: Components for
automation of clinical protocols
Models of protocol concepts
 Programs to plan patient therapy in
accordance with protocol requirements
 Programs to match patients to
potentially applicable protocols and
guidelines

Use of an explicit model to
guide knowledge entry
Model of
protocol concepts
Clinicians
receive expert
advice
CustomKnowledge-base
authors create protocol
tailored
descriptions
protocol-entry
tool
EON
Protocol
knowledge base
Therapyplanning
program
Eligibilitydetermination
program
Model (ontology) of protocol
concepts
Components of the protocol
model (ontology)

Guideline ontology



Defines abstract structure of clinical protocols and
guidelines
Is independent of any medical specialty
Medical-specialty ontology


Defines clinical interventions, patient findings,
and patient problems relevant in a given specialty
Provides primitive concepts used to construct
specialty-specific protocols
An ontology
Provides a domain of discourse for
talking about some application area
 Defines concepts, attributes of concepts,
and relationships among concepts
 Defines constraints on values of
attributes of concepts

Model (ontology) of protocol
concepts
Custom-tailored
protocol-entry tool
Details of CAF chemotherapy
Details of CTX prescription
Custom-tailored
protocol-entry tool: Top level
Specifying eligibility criteria
Use of an explicit model to
guide knowledge entry
Model of
protocol concepts
Clinicians
receive expert
advice
CustomKnowledge-base
authors create protocol
tailored
descriptions
protocol-entry
tool
EON
Protocol
knowledge base
Therapyplanning
program
Eligibilitydetermination
program
Automation of protocol-based
care requires
Ability to deal with complexity of
patient data (e.g., time dependencies,
abstractions, missing data)
 Ability to deal with complexity of
protocol actions (e.g., actions which are
themselves protocols)
 A scalable and maintainable
computational architecture

The EON Architecture
comprises
Problem-solving components that have
task-specific functions (e.g., planning,
classification)
 A central database system for queries of
both

 Primitive
patient data
 Temporal abstractions of patient data

A shared knowledge base of protocols
and general medical concepts
EON is “middleware”

Software components designed for
 incorporation
within other software
systems (e.g., hospital information
systems)
 reuse in different applications of protocolbased care
Components of the
EON architecture
Therapyplanning
component
Clinical
information
system
Eligibilitydetermination
component
RÉSUMÉ
temporalabstraction
system
Chronus
temporal
database
query
system
Tzolkin database mediator
Protocol
knowledge base
Domain
model
Patient
database
Therapy-planning component

Takes as input
 Data
from computer-based patient record
 Knowledge of clinical protocol

Generates as output
 Therapeutic
interventions to make
 Laboratory tests to order
 Time for next patient visit
Episodic skeletal-plan
refinement
1. Flesh out standard plan
from skeletal plan elements
2. Query database for
presence of relevant
patient problems
Protocol
Regimen A
Regimen B
Drug 1
?
Protocol
Drug 2
3. Revise plan based on
problems identified
Regimen B
Drug 1
Drug 2
Domain knowledge derives
from knowledge base
Problem-solving knowledge
automates specific tasks
Domain knowledge
+ Problem-solving method
Intelligent behavior
Problem-solving methods
Are reusable, domain-independent
software components that solve abstract
tasks (e.g., planning, classification,
constraint satisfaction)
 Represent data on which they operate
as a method ontology (model), which
must be mapped to the domain ontology
that characterizes the application area

Mapping domain ontologies
to problem-solving methods
Problem-Solving
Method
Method
Input Ontology
Method
Output Ontology
Domain Ontology
(e.g., clinical protocols)
Problem-solving methods can
automate a variety of tasks

Some skeletal planning tasks




Therapy planning for protocol-based care (EON)
Administration of digoxin in the presence of
possible toxicity (Dig Advisor)
Designing experiments in molecular genetics
(MOLGEN)
Each application entails mapping a
different domain ontology to the same,
reusable problem-solving method
Components of the
EON architecture
Therapyplanning
component
Clinical
information
system
Eligibilitydetermination
component
RÉSUMÉ
temporalabstraction
system
Chronus
temporal
database
query
system
Tzolkin database mediator
Protocol
knowledge base
Domain
ontology
Patient
database
Our goals for eligibility
determination
Automated clinical-trial screening from
institutional and regional databases
 Identification of specific actions that
providers can take to enhance patient
eligibility for guidelines and protocols
 Minimization of inappropriate
enrollment of patients who are not
eligible

EON eligibility-determination
component (Yenta)

Takes as input
 Computer-based
patient record data
 Knowledge of eligibility criteria
of applicable protocols

Generates as output
 List
of patients potentially eligible for
given protocols
 List of protocols for which given patients
potentially are eligible
Classification of eligibility
criteria for clinical trials
Stable (e.g., having received prior
therapy)
 Variable (e.g., routine lab data)
 Controllable (e.g., use of a given drug)
 Subjective (e.g., likelihood of
compliance)
 Special (e.g., lab data requiring invasive
or expensive tests)

Qualitative eligibility scores
For each eligibility criterion,
for each point in time,
the computer assigns a score:





P
PP
N
FP
F
meets the criterion
probably meets the criterion
no assumption can be made
probably fails the criterion
fails the criterion
Eligibility criteria derive from
the electronic knowledge base
Use of an explicit model to
guide knowledge entry
Model of
protocol concepts
Clinicians
receive expert
advice
CustomKnowledge-base
authors create protocol
tailored
descriptions
protocol-entry
tool
EON
Protocol
knowledge base
Therapyplanning
program
Eligibilitydetermination
program
Components of the
EON architecture
Therapyplanning
component
Clinical
information
system
Eligibilitydetermination
component
RÉSUMÉ
temporalabstraction
system
Chronus
temporal
database
query
system
Tzolkin database mediator
Protocol
knowledge base
Domain
model
Patient
database
Tzolkin database mediator
Serves as a common conduit for all
problem solvers that must access
patient data
 Embodies components that address
significant problems in temporal
reasoning

 RÉSUMÉ—Temporal
abstraction
 Chronus—Data query and manipulation
RÉSUMÉ
temporal-abstraction method
 Takes
as input primary patient data
and previously determined
abstractions of those data
 Generates as output further
abstractions of the input
 Requires a separate knowledge
base of clinical parameters and
their properties
The temporal-abstraction task
PAZ protocol
BM T
Expected CGVHD
M[0]
Platelet
counts
(• )
²
•
150K
²
•
²
•
²
•
² ²
• •
²
•
M[1] M[2] M[3]
²
•
²
²
•
•
100K
0
50
100
²
•
²
M[0]
²
²
•
200
Time (days)
.
•
•
²
•
M[0]
²
•
²
•
Granuloc yte
counts
²
(² )
•
2000
1000
400
Knowledge required for
temporal abstraction

Structural knowledge
(e.g., definitional relationships among lab tests and clinical
states)

Classification knowledge
(e.g., how numeric values map into qualitative ranges)

Temporal-semantic knowledge
(e.g., whether intervals are concatenable or downward
heriditary)

Temporal-dynamic knowledge
(e.g., minimal values for a significant change, functions to
predict persistence of a value over time)
Acquiring temporal-abstraction
knowledge for RÉSUMÉ
Model of
clinical parameters
Tool for entry
of temporalabstraction
knowledge
TZOLKIN
Abstractions
of relevant
clinical
parameters
Knowledge-base
authors enter knowledge
required for temporal
abstraction
Parameter
knowledge base
RÉSUMÉ
temporalabstraction system
The EON Architecture
Problem-solving components that have
task-specific functions
 A central database system for queries of
both

 Primitive
patient data
 Temporal abstractions of patient data

A shared knowledge base of protocols
and general medical concepts
A protocol model shared
among all components
Makes explicit relevant assumptions
about the application domain—we know
what our programs know
 Consolidates the task of maintaining
the domain knowledge—all the
knowledge is in one place and can be
examined in a coherent fashion

Planned applications of EON
Hypertension guidelines at Palo Alto
VA Health Care System
 Fast Track Systems, Inc., plans to
develop systems for automation of
clinical trials

EON’s component-based
approach allows
Developers to create new problemsolving modules that “plug and play”
 Clinicians to create new guideline
knowledge bases that can interoperate
immediately with existing components
 System architects to integrate
components with other software
modules using standard
communication methods

Some implications of our work
Enhanced authoring, maintenance, and
execution of clinical protocols and
guidelines
 Incorporation of guideline-based
practice into routine patient care
 Increased participation of communitybased practitioners in clinical research
