2. Capturing and Documenting expertise is critical to its evaluation and
verification, sharing, and preservation
Semantic Networks are useful to document a process or object or topic and
visualize its relationships and characteristics — an aid to modifications
and improvements
Decisions Trees are an effective way to document the decision-making
process, but are limited to problems with a pre-determined set of possible
solutions — such as possible causes of failure for a part or a list of
treatments for a known medical condition
Something more flexible than semantic networks and decision trees is required
to address the dynamic problems and solutions found in research…
p
3. A hybrid Expert Decision-Support System targeting Melanoma
could/should/would-
• Capture and share department staff expertise to enable each to use their
work and research time the most effectively
• integrate patient(s) results from laboratory tests and instrumentation
• integrate patient(s) medical imagery analysis
• serve as model expert system architecture for other diseases and
conditions
diti
• support practical and effective remote access to Mayo resources and
expertise (telehealthcare)
4. Premise — human beings have arrived at the top of their food chain,
presumably because they solve problems the most effectively — therefore,
how humans solve problems, their “intelligence”, is worth emulating*
Artificial Intelligence (AI) can then be defined as the capturing of human
p
problem-solving capability into machines, in either dedicated hardware or
p
reprogrammable software
Attempting to capture all of the integrated and generalized problem-solving
p g p g g p g
intelligence of a human being is a gigantic challenge**, but replicating that
intelligence for a specific problem is very achievable
5. An Expert System is a reasonable solution to the classic AI problem of
programming [human] intelligence into a machine - it mimics the decision-
making process(es) of a human being
• This can be accomplished through heuristics, pattern-recognition, inexact
logic, or visual image analysis (as well as tables, hand computers, and
other devices)
• It can be designed and approximated using flow charts and decision trees,
but should be implemented on a computer for speed and flexibility
• To be effective, ideal expert systems are intentionally focused, not broad
nor too all-inclusive
• They have a pre-planned mechanism for learning and improving
Th h l d h i f l i di i
• They provide detailed explanation of the arrived at decision. The
explanation portion of the expert system is commonly referred to as a
Decision Support System and is often more desirable than pure machine
pp p
decision-making
6. Facts can be in the form of text words,
Knowledge
instrument measurements (including Base
time-based), and/or imagery Knowledge Base cab be built from Heuristic
Rules, Instrument Measurement Exemplars,
Imagery Libraries
Facts
Expertise Inference
Engine
A variety of Inference Engines are
y g
Expertise can be in the form of text available, including
(instructional words), values, and/or imagery Production Systems,
Neural Networks, Fuzzy Arithmetic, and
Statistics
Expert System
7. • Permanence - Th d l
P The declarative interpretation and preservation of tacit
i i i d i f i
personnel knowledge, preserving expertise of human expert(s) and
protecting the investment in their education
• Increased Availability and Speed- free-up senior domain expert for more
y p p p
complex problems, available around-the-clock, fast
• Reduced Cost by shifting the burden of routine decision-making, more
effectively using senior staff for difficult probelms
• Combined Experience and Expertise of multiple team experts in efficient
manner
• Education Aid - Explanation (for training, experiment planning, and liability
defense)
• Increased Reliability - steady, unemotional, and complete response at all
times
• Error Reduction - In a rapidly advancing field such as [medical] research,
capturing the latest procedures and treatment steps in an expert system
minimizes errors and aids entire team in moving forward in unison
8. • Successive Application of Heuristics (IF-THEN Production Rules),
• Numerical Pattern-Recognition,
• Inexact Logic [Fuzzy Classification
and decision-making], or
• Imagery Analysis with
Feature Identification
When implemented, each strategy
provides a User Interface and
some manner of Reporting results,
but each approach infers knowledge from
the facts/data differently. A few
examples will help illustrate this…
Common form-based user interface
for rule-based expert system
9. A Production System is the most familiar type of expert system, with the following common components -
• p
A Collection of Rules or Knowledge Base linguistically describing the problem and situation, each with a set
of requirements for some condition to be met (the antecedent) — triggering an action or actions to be
taken (the consequent)
• An Inference Engine to implement abductive (‘guessing’ first to reduce search space), deductive (fact-
driven) and inductive (goal-driven) reasoning by sequentially and repeatedly evaluating the prioritized rules
using supplied facts - producing new facts
i li d f t d i f t
• If the antecedent fact(s) are known, then the consequent fact(s) can be derived or inferred and thus
knowledge is gained
• An Explanation Facility to produce a record of how the reasoning was achieved
• AK Knowledge Acquisition F ilit — t provide an automated way to enter knowledge without requiring
l d A i iti Facility to id t t d t t k l d ith t ii
explicit coding
Example- (repeated for each mole mapped)
IF: Working Memory
Knowledge Base Inference Engine (patient or group
The patient … has [mole_n.changing] =yes
p [ _ g g] y facts)
(heuristic l )
(h i i rules)
AND: The patient … has [mole_n.shape] = odd
AND: The patient … has [mole_n.color] = black Agenda
THEN:
[Biopsy mole_n]
Explanation Knowledge
[Send sample mole2 to pathologist #1] Facility
F ilit Acquisition
[Send sample mole2 to pathologist #2] Facility
Confidence = 10
10. Pattern-recognition examines a set of data and either produces a value/values (such as a dosage), or, classifies
the sample (categorize the patient’s condition). Popular classifiers include:
• k-Nearest Neighbor (k-NN), and
• Margin Classifiers (important for medical use, includes Support Vector Machines.
• Soft-Margin Classifiers (border on Fuzzy Classification),
• Neural Net classifiers (radial basis function and probabilistic neural networks).
Most approaches involve the following stages
• Coding a series of [patient] measurements and representing them as a vector of features (Feature Vector)
[Sex Age HT WT T SBP /DBP LungCap Cholesterol Hemoglobin]
[M 47 72 208 98.6 118/78 6 190 14.5 ]
• Evaluating a p
g pre-calibrated transfer function using the Feature Vector as inputs to generate a desired
g p g
value (i.e. duration of treatment)
a(Sex) + b(Age) + c(HT) + d(WT) + e(T) + f(SBP) + g(DBP) + h(LungCap) + i(Cholesterol) + j(Hemoglobin) => patient dosage
Or
• Comparing new “unknown” feature vector against others in a library of “knowns” using numerical techniques
p g g y g q
until an acceptable match is located (this is referred to as Supervised Classification).
• Classify sets of new feature vectors into categories (also called classes) by grouping like vectors into
clusters or natural classes (Unsupervised Classification)
All pattern recognitions approaches can produce confidence measures (CF’s) to indicate how well an unknown
was assigned to a category
11. –
Patterns from collections
P tt f ll ti
[a
of point measurements
b
Feature Vector = [a b c … n]
c
…
n]
Patterns gathered from strip-chart type data
Feature Vector =[a(t) a(t+1) a(t+2) … a(t+n)]
ue
riable a valu
Var
Time (t)
12. –
[M 47 72 208 98.6 118/78 6 190 16.5]
Euclidean distances
Category exemplar vectors and tolerance value
g y p
[M 42 72 208 98.6 118/78 6 190 16.5]
[F 47 72 208 98.6 118/78 6 190 13.5]
[M 35 72 208 98.6 118/78 6 190 14.1]
[M 30 72 208 98.6 118/78 6 190 15.5]
Unknown Vector so different
it requires new category
i t
[F 28 72 208 98.6 118/78 6 190 14.2]
[M 24 72 208 98.6 118/78 6 190 15.2]
98 6 15 2]
[F 22 72 208 98.6 118/78 6 190 14.5]
Unknown Vector within acceptance
tolerance of existing category
[M 24 71 206 98.8 117/79 6 188 14.9]
[F 18 72 208 98.6 118/78 6 190 13.52
13. Fuzzy Logic for decision-making is especially effective in complex system where
precision and significance become mutually exclusive. Even when ambiguity and only
incomplete information is available - decisions are still required to be made
made.
Inexact but weighted strategies help launch decision-making when little initial data
are available
• Problem Measurements or
Condition Variables are mapped into
a Fuzzy Membership Set
• Linguistic descriptions
(heavy d
(h dose, high temperature,
hi h
low blood pressure, dark rash, etc) Boundaries + Core = Support
help define the Boundaries and Core of a fuzzy variable membership set
• Linguistic operators called hedges modify the geometry of a membership set using
a mathematical formula (“Very” a = a2, “Slightly” a = √a., etc)
• Fuzzy variable membership sets are processed using Fuzzy
Arithmetic/Logic/Calculus, producing Result Sets which are then Defuzzified (by
p
centroid or other methods) into Actionable Values (dosage, duration, etc)
14. • Fuzzy Logic provides a mechanism for reasoning despite the presence of
uncertainty. Fuzziness describes and manages ‘ambiguity’ - a different type of
uncertainty than random chance or [conditional] probabilities Ambiguity involves
probabilities.
possibilities and belief, a measurement of information and knowledge, not a
statistic
• Fuzzy not suitable where precision is a requirement, but, experimental medicine
y p q p
doesn’t require this, i.e. dosages, durations of treatment, combinations, etc
• Fuzzy and Rule-based decision-making are often combined
IF patienttemp is normal THEN aspirindose is none
IF patienttemp is low THEN aspirindose is none
IF patienttemp is high THEN aspirindose is small
IF patienttemp is very high THEN aspirindose is large
IF patientpain is typical THEN asprindose is average
IF patientpain is severe THEN asprindose is extra
If p
patienttemp is high and patientpain is typical, the aspirindose is both small and
p p p p p
average, and they are geometrically combined and the new centroid is computed
(Defuzzification) to determine the final dosage
15. Imagery Analysis is itself a specialized type of expert system
• Images coded into intensity or color values, and stacked into layers
• The same pattern-recognition and classification used for vectors can be
expanded and applied to 2-D, 3-D, multi-dimensional, and hyper-
p pp , , , yp
dimensional data (X-ray, CT and CAT, PET, MRI)
• Results are typically in the form of imagery as well — supporting easy
human interpretation
p
Patterns gathered from imagery type data
550604050
555565004
Feature Vector = [ 550005050
500550870
]
500550880
050007 570
17. (216) different Medical Practice Management software systems available
(recent Capterra search Aug2009)
Few actual Medical Diagnostic Applications — mainly because of liability
issues. The few that ‘are’ out there are literally plastered with disclaimer
statements
This is strong defense that medical expert systems are best used internally
by physicians not potential patients If currently available diagnostic
physicians, patients.
knowledge is captured and used, then it will be available in the future when
legal issues are resolved
18. Professor Ed d Feigenbaum — E
P f Edward F i nb m Expert systems, C n i M ll n University
t t m Carnegie Mellon Uni it
Professor Edward H. Shortliffe — MYCIN, Standford University
Professor Charles Forgy — Rete scheduling algorithm, Carnegie Mellon
University
Professor Lotfi Zadeh — Fuzzy logic, University of California, Berkeley
Professor Martha Evens — CIRCSim-Tutor, IIT, Northwestern University
Professor Harry Pople — Expert Systems (of the University of Pittsburgh)
Dr. Jack Meyers — Medical diagnostician, University of Pittsburgh
19. • DENDRAL, 1960’s
• MYCIN, 1970’s
• INTERNIST — Diagnose Internal Medicine problems, 1975
• MOLGEN — Plans experiments in molecular genetics, 1977
p g
• CASNET — Medical Decision-making for Glaucoma Treatment, 1978
• DENDRAL — Identify Organics by Mass-Spec Signatures, 1978
• CENTAUR — Analysis of Pulmonary Tests, 1979
Tests
• EMYCIN — (empty MYCIN), 1980
• NEOMYCIN — Another derivative of MYCIN, 1980’s
• CADUCEUS — Diagnosis of blood-borne disease, mid-1980’s
• CIRCSim-Tutor — Teaches cardiovascular medicine, 1993
• Easy Diagnosis (http://easydiagnosis.com/ ), MatheMEDics, 2008
g ( p g )
• Medical Expert On-Line (http://med-expert.net/ ), 2009
20. There are many different clinical tasks to which Medical Expert Systems can
be applied:
• Alerts and Reminders - Real-time warning of a patient’s condition
• Diagnostic Assistance - Formulation of likely diagnoses
• Therapy and Treatment Critiquing and Planning - Critiquing systems can
look for inconsistencies, errors and omissions in a treatment plan
• PPrescription Decision Support Systems - P
i i D i i S S Prescription of medications and
i i f di i d
can assist checking for drug-drug interactions and dosage errors
• Information Retrieval - Assist in formulating appropriately specific and
accurate clinicall queries
li i i
• Image Recognition and Interpretation - X-rays, angiograms, CT, CAT, PET,
and MRI scans can be automatically interpreted flagging potentially
abnormall images for detailed human attention
b i f d il d h i
21. • Anecdotal record-keeping (patient “charts” often contain physician notes
more so than tabulated values)
• Disparate records from multiple physicians/hospitals
• Diagnoses based upon significantly differing backgrounds and experiences
g p g y g g p
• Non-standardized medical (clinic or hospital) patient information systems
• Staff technophobia and limited time for learning new systems
22. • Liability issues — who’s at fault if errors are encountered? If an expert system
who s
produces an incorrect solution when used to help you diagnose your vacuum
cleaner — no one cares. But if an expert system is used to diagnose a patient
and it errors — everyone cares, patients, original domain experts, and software
developers alike
• Medical (life and death) expert systems suffer from an application-specific
phenomena- incomplete exemplars. (If a patient recovers — they often cannot
be completely inspected (dissected) to see “what happened”, and even
p p pp
patients who do die don’t always allow the body to be examined to evaluate the
applied solution.) This leads to incomplete and even corrupted learning by any
human expert or system intended to emulate that human expert
• Shortage of industry/field-wide standards for information exchange
industry/field wide
These barriers led the away from the adoption of full-authority expert systems to
something more supportive to aid human decision-making…
g pp g
23. "Clinical D i i n S
"Clini l Decision Support Systems link health observations with health
tS t m h lth b ti n ith h lth
knowledge to influence health choices by clinicians for improved health
care"
• Decision Support Systems often can provide many “acceptable” solutions
acceptable solutions;
their true value is then in guiding the physician away from “poor” solutions
and disasters. This also has been shown to be the more efficient and
resource economical
resource-economical strategy as well Much much more time and effort
well.
can be spent finding and ideal or optimal solution when the better use of
the resources would be toward avoiding anti-optimal solutions and just
p
providing ranked acceptable choices
g p
24. Medical Expert System and Clinical Decision Support System design aided by
shared standards for information
Health Level 7 (HL7)
• HL7 creates international standards (i.e. the Arden syntax) for inter-
( y )
system and inter-organization messaging, for decision support, clinical
text document mark-up, user interface integration as well as a health data
model and a message development methodology
• "Level 7" refers to the top level (the application layer) of the seven-layer
International Standards Organization's (ISO) communications model for
Open Systems Interconnection (OSI). The application level addresses
definition of the data to be exchanged
• HL7 has 4,400 members, has achieved ISO status for its standards, and
is moving toward harmonization with the European CEN standards
25. UMLS Semantic Networks
• The UMLS Semantic Network is one of three UMLS Knowledge Sources
developed as part of the Unified Medical Language System (UMLS)
project. The network provides a consistent categorization of all concepts
represented in the UMLS Metathesaurus
26. p y pp y
Medical Expert Systems and Clinical Decision Support Systems can playp y
important roles in Evidence-based Medicine (EBM)
• Applies evidence gained from the scientific method to certain parts of
medical practice. It seeks to assess the quality of evidence relevant to
p q y
the risks and benefits of treatments (including lack of treatment)
• EBM used techniques from science, engineering, and statistics, such as
meta-analysis of medical literature, risk-benefit analysis, and randomized
y y
controlled trials (RCTs)
• The decision-reporting capability of an expert or decision-support system
p
produces the proper evidence required for effective EBM
p p q
27. • Medical Patient Information is gathered and stored in a variety of forms
(measurements, charts, imagery, etc.), therefore a Hybrid Expert System
should combine all of these sources to perform the best possible decision
Knowledge Base
making.
Hybrid Expert System
• Expert systems Heuristic Rules
should be
designed to aid Integration Engine
Instrument
planning Pattern
Working Memory (patient
or group facts)
g p )
experiments
i Recognition
R
Coordination
and clinical Strategy
studies Image Analysis
• Hybrid E
H b id Expert System
S
Knowledge
Explanation
would keep department Facility
Acquisition
Facility
knowledge the most
current and promote their research
to the leading edge in its field
28. • Capturing and documenting gained
knowledge using semantic network
diagrams — as an aid to analysis
and planning
• Case Study Dissection and
anecdotal record inspection
and declarative reduction into
data (facts)
• Implement E-mail and SMS
Text messaging input and
output from Medical Expert
System
• Support partial input of needed rule
facts (data) followed by piece-meal
output availability
29. • Patient treatment history
• Patient
P ti t response hi t
history
• Clinical study results
• Heuristic rules as the back-end interface combining varied data
• Data-mining instrument responses to provide valuable inputs to rules
• Introduction of instrument measurements and trends into rule-based
system, including medical-tailored data-logger
IF (feature x)
AND NOT (feature y)
THEN (action b)
Pattern-
Knowledge Base
Recognition (heuristic rules)
Engine feature x,
feature y,
feature z
30. An Expert Decision Support System tailored for the research and treatment
of melanoma would adopt the features and strategies for Medical Expert
Systems described above, but should also explore refinements and
characteristics unique to its own topic
31. • Document treatments applied (as a series of heuristics), and track
performance (as facts). This serves the preservation task
• Analyze clinical trials results, and plan next studies by watching for
patterns in patient treatment and response across individuals and larger
numbers of patients - treatments and responses are patterns for
predicting performance
• Accept results from case studies and predict future performance of
treatments
32. • Decision report to provide (doctor or researcher) source information about
each rule, and confidence f t attached t each s b din t d isi n
h l nd nfid n factor tt h d to h subordinate-decision
• Heuristics regarding treatment to be physician-specific, satisfying multiple
rules for the same facts thereby reinforcing confidence for decisions when
multiple physicians have same experience on a subject, and attenuating
p p p
confidence where they have differing experience
• Decision report to be generated as check-list and ‘checked’ by human
expert contributors for accuracy and appropriateness, saving significant
amounts of time, and p
, providing documented paper-trail for insurance
g p p
claims and defense during legal actions
• System decisions not different from human experts, but more complete,
and arrived at much more quickly
• Specialized reports to be prepared to help analyze clinical trial results
results,
plan improvements, and select new directions.
33. • Architecture to utilize commercially-supported development environment
• Operate using integrated Mayo Melanoma-specific modules, interfacing to
Mayo Medical Informatics
• Server-based, in parallel to existing informatics but not integrated — no
, p g g
downtime for existing informatics, only new reports to be defined
34. Phase 0 — Feasibility Study
Phase I — Heuristic Infrastructure
Phase II — Pattern-Recognition Extension Module
Phase II — Image Analysis Extension Module
Phase IV — Expansion of System, Licensing,
Re-use
Re use of Architecture
35. • Analysis of Existing Knowledge (staff experience, data sources, formats,
generation, conversion, support, and maintenance)
• Design Architecture and Draft Plan
• Model one sub-type of melanoma and a subset of its treatments and
yp
feedback first, then scale up this complete start-to-finish path to include
more, use first example to accurately estimate cost, time, and effort
• Complete Verification and Validation testing to insure accuracy and
p g y
repeatability
• Prepare federal R&D proposals — NIH Small Business Innovation Research
(
(SBIR)/Small business Technology Transfer Research (STTR) program —
) gy ( )p g
“Expert Decision Support System for Melanoma Research and Treatment”,
Small Business — Vogtland I&R, medical facility partner — Mayo Clinic,
explore women & minority contribution opportunities
36. –
• Dissection and Declarative Coding of Existing Patient Records — Creation
of First Production System Heuristics
• Interviewing [Interrogation] of Melanoma Domain Experts — Refinement of
Production System Heuristics
• Coding Production System Heuristics into Version 1.0 Web-based
Melanoma Research and Knowledge Base
Treatment Expert System Heuristic Hybrid Expert System
Rules
R l
• Complete Verification and Integration Engine Working Memory
Validation testing to insure (patient or group
facts)
accuracy and repeatability Coordination
Strategy
• Licensing and Subscription
Distribution Explanation
Knowledge
Acquisition
Facility
Facility
y
37. –
• Integration of Pattern Recognition for
Patient & Study Group Symptoms -> Condition or Disease
• Produce PatternRec Extension Module [Version 2.0] Web-based Melanoma
Research and Treatment Expert System
• Complete Verification and Validation testing to insure accuracy and
repeatability Knowledge Base
Hybrid Expert System
• Licensing and Subscription
g p Heuristic
Rules
R l
Distribution Integration Engine
Instrument Working Memory
Pattern (patient or group
Recognition Coordination facts)
Strategy
Knowledge
Explanation
Acquisition
Facility
Facility
y
38. –
• Integration of Imagery Analysis for PET, CT, MRI, X-Ray, etc
• Produce ImageAnalyst Extension Module [Version 3.0] Web-based
Melanoma Research and Treatment Expert System
• Complete Verification and Validation testing to insure accuracy and
p g y
repeatability
• Licensing and Subscription Knowledge Base
Hybrid Expert System
Distribution Heuristic
Rules
R l
Instrument Integration Engine Working Memory
Pattern (patient or group
Recognition Coordination facts)
Strategy
Image
Analysis
Knowledge
Explanation
Acquisition
Facility
Facility
y
39. –
• Refinement of Integrated Pattern-Rec + Image-Analysis + Production
y
System Heuristics
• Introduction of Hybrid Self-Learning (pattern-rec and imagery into
Heuristics)
• Complete Verification and Validation testing to insure accuracy and
repeatability
• Licensing and Subscription Lab data Rules
Distribution
• Re-Use of Melanoma
Architecture for
Additional Diseases
Melanoma Lymphoma Clinical data Validation and Approval
Leukemia
L k i
Day-to-day operations
40. Once up and performing, melanoma expert system can be self-supporting by
Licensing and Subscription fees — allowing addition to and use by other
g p g y
hospitals and clinics
Melanoma-specific
Knowledge
Either content or system
architecture can be
licensed
Non-melanoma topic
License
Internet
Subscribing clinics and Manager
research programs
Originating Mayo
Contributors pay department users
reduced user fee
41. Initial melanoma system will become a model for additional systems for other
diseases and physical conditions, leading to a finalized matrix expert
system that can cross domains when appropriate with simple rules to
combine rules from different fields.
Melanoma Lymphoma Leukemia
License
Internet
Subscribing clinics and Manager
research programs
Originating Mayo
Contributors pay department users
reduced user fee
42. “ ”
Important concept — the Expert System is always improving — knowledge is
always being gained. The addition of new rules and classifications adds to
the knowledge base and reinforces the existing rules increasing their
confidence. Failures and errors provide heuristics for what did not work
and under what conditions it didn’t work. This is very important- the
concept of providing information of what not to do as well as what to do.
ty
& Capabilit
Both have value. Additional Domain
Introduction of Imagery Expert Interrogation
Analysis Self-learning through
use and new data
Usability &
Introduction of Pattern
Self-learning through
Recognition
use and new data
Initial Dissection Self-learning through
of Medical
of Medical use and new data
Records into
Heuristics
Self-learning through
and …
use and new data
…(initial) Domain
…(initial) Domain
Expert Interrogation
Effort & Time (measured in system usage)
43. The expertise, software tools, and components to complete this project all
exist, but this work needs to be initiated -to preserve and share the
knowledge gained by Mayo’s experience in treating melanoma, and in
treating other diseases
45. ’
Expert Decision-Support System for
p pp
Hydro-pneumatic Suspension Unit
UML
(HSU) design and operation
Deployment
Diagram
Screen captures of portions of
p p
the HSU and EFV Decision-
Support System User Interface