An Expert System Interpreter for Time Course Data
with Refinement in Context
From: AAAI Technical Report SS-94-01. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved.
P. Preston, P. Compton,D. Litkouhi
Schoolof ComputerScience and Engineering, University of NewSouth Wales
Sydney2033 Australia (pkilp@cse.unsw.edu.au,compton@cse.unsw.edu.au,
dar@cse.unsw.edu.au)
G. Edwards
Departmentof ChemicalPathology,St Vincent’sHospital
Sydney2010Australia (glenn@cse.unsw.edu.au)
Introduction
Knowledge
acquisition and knowledgemaintenanceare
problemswith any expert system. Theseproblemsare
exacerbatedin domainsdealing with temporaldata such
as the exampledata sets distributed for AIM-94.
Knowledgeacquisition for such domains requires
informationabouthowfeatures In the data are identified
as well as howthese features are reasoned about.
Ripple DownRules (RDR)Is a knowledgeacquisition
methodologywhich goes somewaytowards addressing
these problems.
The problem with knowledge acquisition
and
maintenanceis that the knowledgean expert provides,
whetheraboutfeature Identification or reasoningabout
features, cannot be guaranteedto apply outside the
context in which it was provided by the expert
[Compton,1990]. The RDRapproach attempts to use
the knowledge
only In the samecontext in whichit was
provided. For rule based systems a new rule (a
correction) ts attached at the end of the sequenceof
rules which had been evaluated giving the wrong
conclusion.This rule is evaluatedonly after the same
rules are evaluated with the same outcomes as
previously. Rulesare never removedor corrected, only
added. All rules providea conclusion, but the final
output of the systemcomesfromthe last rule that was
satisfied by the data. The case that promptedthe
addition of the rule Is also stored to validate further
knowledge
acquisition.
Onlyvalid rules are allowed.Avalid rule is one that
will correctly interpret the newcase for whichit is
addedbut will not be satisfied by the case associated
with the previouslast true rule in the sequenceof rules
evaluated. The expert Is required to choose rule
conditions from a list of differences between the
currentcaseandthe caseattachedto the last true rule to
ensurea valid rule.
It might be expectedwith the RDRapproachthat rule
contexts would be too specific resulting in much
repetition and redundancyin the knowledge
base. This
does not occur [Mansurl, 1991; Gaines, 1992]
presumablybecausethe approachallows the expert to
focus on domain Issues rather than knowledge
engineering.Amajorvalidation of the approachis the
PEIRSsystemin routine use tn the Dept. of Chemical
115
Pathology, St.Vincent’s Hospital Sydney[Compton,
1992, Edwards1993]. PEIRSnowhas morethan 2000
rules. It covers about 25%of ChemicalPathologyand
is over 95%accurate. Thesystemwasput into routine
use after 200 rules were added. All other rules have
been added by experts as errors occur, without
knowledge
engIneeringassistance. It takes about three
minutesto adda newrule andthe laboratoryanticipates
continuouson goingdevelopment
as it is trivial to add
to the systemas domainknowledgeevolves.
Thedata used by PEIRSis characterised by multiple
test results on sequential specimenssubmittedto the
diagnostic laboratory.
The major knowledge
engineering effort with PEIRSwas deciding what
functions had to be applied to the data and what
features had to be extracted. Nine fairly simple
functions such as MIN, MAX,and AVERAGE.
Transformations of data allow abstractions such
CURR(BLOOD_C02)
is HIGHto be used in rules.
Thesefeatures and functions are hard coded. This very
simple approach only works because the features
identified can be refined in context. That is, a rule
sequence may include TSHis HIGHand TSH<10.
Simplemathematicalexpressions can also be included
in context [Srinivasan 1991]. Thesenewfeatures and
functions are not namedand so cannotbe used in other
contexts, but they allowthe systemto deal with fairly
complextime course data with very simple predefined
features andfunctions.
Thediabetes data in AIM-94
data is likely to require
other features and functions which are even more
difficult for the expert to pre-identifythan with PEIRS.
A newimplementation of RDRhas been developed to
address this problem. TCRDR
(Time Course Ripple
Down
Rules) uses a moreflexible interpreter, allowing
for the creation of reusabledata abstractions. Aswell,
the user interface has been de-coupled from the
Inferenceengine, allowingfor greater portability - the
inference engine does not need to be on the same
machineas the interface. It also allows for easy
eustomisationof the interface, as wellas testing of the
interface and inferenceengine.
In TCRDR,
internal data is essentially stored in a 2
dimensionalarray, very like the appearanceof the raw
data. Withthis representationit is possible to directly
address individual elementsof the array. It is also
possible to reference data points temporally, as in
referencingthe valueof a test at sometimein the past.
A set of operators is supported that allows
combinationsof data references to be combinedinto a
functionor feature. Thesefirst order abstractions can
also be combinedto form higher level features and
functions, leading to arbitrarily complexfeatures. New
features and functions can be addedat any time in the
systems development.Howeverthe central feature of
the RDRapproachremainsin whichrules are addedin
context allowingfeatures to be refined in context. As
with PEIRS,this allows simpleand inadequatefeatures
and functions in context as required. Theburden of
trying to achieve perfect feature identification is
removed by allowing local fixes to inadequate
definitions.
definitions as requiredandadd rules. This is a slightly
moredifficult task than just addingrules as the expert
had to understand the function interpreter syntax.
However, no knowledge engineering knowledge is
required becausethe RDRapproachavoids the problem
of havingto understandpotential interactions between
rules.
Someknowledgeengineering/programmingassistance
is still required in setting up and appropriate data
representation. Themain environmentdevelopedfor
the expert in workingwith the data can be seen in Fig
1. Theexpert wasinterested in reviewingthe data in
the form of sliding windowsat different times. Data
typically werenot dense, so that a window
widthof 11
wouldprovide several days data at once. Internally
provision wasmadefor workingwith windowsof some
40 columnsmaximum.
It mayseema restriction but as
it is a sliding window,
data can be reviewedfor periods
in the past at anytime.
Aim
Theaim of the present study wasto build a systemto
deal with the AIM-94
diabetes data in whichthe expert
was expected to add his ownfunction and feature
RDR Shell/AIM
Figure1
HI~
W gls,~oll.
Inb=14~l tcmlght’S
I,umlln
! O =I
](PlUI~II0,1O0)
~.
iiGet Windowl
i" I(w’~.
t
LestTruelIs ILesl RuleI 9~II mg~mol
II moDmg
Ilnun eutollSu
I
IITi .....~.ao12.~m.oon.oou.oot2.aai..oon.~ ~.oom.oo
Imrll
....
0.00
io
i lU~.g,,.;~
....
~p~t-onm
0.17
4
0.42 O.SI 1.00
io
141 10
I.I?
4
............................................
II
1.42
I0
I,M 2.~
14
~I
2.~
~ 10
r"li
~n
IEI
IF’"
................"............"........ |1
::::::::’"::::::::::::::::
mn
|1
............................................
|1
H i
gi~d
|!
............................................
The main knowledge
acquisition screen. The
experthas facilities hereto
review windowsof data,
depending
on how
temporallythe rules he Is
requiredto make,test data,
call up screens for making
new rules or reviewing
existing
rules
or
conclusions.
II............................................
|1
T~lp.[xi;
¯
Li~..F.ml
I~ ~,~nt
...o
IleO-s~ ....
¯ leO.~ ....
I
o ......
.
.
....
o ......
............................................
............................................
o.z? o.,~ Ill.rl
3.50 9.49
1.2"/ 6.44 Ill.r/
1.110 ll.il4
.
...................
1.21 6.~I "/.22
12.21
?.~13 ,I.61
S.:N
9.99
6.94
10.06
Systemaim
Thediabetes data wasused to build rules designedto
assist patients monitortheir insulin therapy for the
followingreasons[Hauser,Kraegenet el. 1992]
i) most patients monitoringtheir ownglucose levels
are highly motivatedand generally awareof the
purposeof monitoring.
ii) patients, In concertwiththeir physicians,develop
a goodsense of their individual responses to
insulin therapy and to environmental factors
which impact upon glucose levels and insulin
requirements.
nltl
I~I
~;),lI
iv)
v)
g
for the support of family or communitycaters
whomaybe required to supervise monitoringbut
have less sophisticated knowledge of the
principles of monitoring;
to assist clinicians and diabetes specialists in
periodicassessmentsof patients’ cumulativedata.
Wedo not believe these systemsshould replace or in
any way subsumethe patient’s role in monitoring
his/her diabetes.
Thefollowingtypes of output wereseen as potentially
useful and wereincorporatedinto interpretations
generated by the RDRtool built for the AIM’94
dataset:
Themostappropriateuse of AItechniquesis then:
as a "back-up" system to alert patients to t)
indicationof undesirablyhigh or lowresults;
i)
importantdata that mayrequire intervention;
ii) recommendations
for the nature of intervention;
ii) to critique the appropriatenessof interventions;
iii) recommendations
for the timing of intervention.
lii) as an educational tool for patients entering a
monitoringprogram;
116
Wedid not attempt to model individual’s responses to
insulin therapy or provide quantitative estimates of
adjustmentsto insulin therapy.
The following are simple examples of refinement.
In clinical
practice,
a number of factors may
significantly alter our approach to managementof
diabetes. These include the patients age, presence of
co-existing Illness, other drug therapy and general level
of competencein self-monitoring. The absence of this
data forces us to assume all patients are "average"
patients - a tenuous concept which constrains our
systems’ ability to fine-tune its recommendations.
Preg-BG > 180
=> High fasting glucose.
Considerincreasing night-time Insulin.
Thus, the following clinical
principles
were
incorporated into rules:
i)
aim to maintain glucose levels between 80 - 180
mg/dL;
ii) modifications to Insulin dosage were suggested if
sustained high (or low) levels were obtained over
2-3 days. To this end, functions for calculating 2day and 3-day meanswere written.
iii) more Immediate intervention was recommended
for more extreme high or low values (< 60 or
280 mg/dL)
Results
The expert added a numberof functions to the system,
including:
CURR(x) :x[0]
PREVI(x)¯ x[-1]
PREV2(x): x[-21
most recent value of a test
2rid most recent value
3rd most recent
: prev2[x] %prevl(x) %curt(x)
IMIN(x)
MAX(x): prev2[x] $ prevl(x) $ cuff(x)
(% = minimumoperator; $ = maximumoperator)
In the context of AIM-94diabetes, the definitions of
MAXand MINwould allow for finding the MAXor
M/NBGmeasurementsfor a particular period - such as
the maximumpre-breakfast BG
Other definitions Include
last_two_average(x):
(CURR(x) + PREVI(x))/3)
last_three_average(x)
(CURR(x) + PREVI(x) + PREV2(x))/3)
I
which simply averages the last
measurementsfor a reading.
three
available
At the time of writing, the experthad added 63 ripple
downrules whichused the functions above to deal with
the AIMdata set. Further rules are being added, and
additional functions maybe required. The key element
in this approachis that the very simple global functions
above can be refined in context with the RDR.At this
stage the expert does not have to consider the global
utility of the refinements as they only apply to the
context, the particular chain of rules.
117
This rule was entered
I
It was then discovered that it didn’t apply in the
following circvmstances
Prevl (preb_BG) <180
=> High fasting glucose. Yesterday’s level OK.
Suggest: let’s see fasting level tomorrowbefore
changing anything.
I
Another Independent refinement was also added
Prevl(preb_BG) >180 and
Prev2(preb_BG) < 180
=> High fasting glucose over consecutive days.
Try increasing night-time insulin dose tonight.
I
In the tree built so far, the longest chain of
corrections before reaching a final conclusion was
five.
Discussion
With PEIRS, we showed that the RDRapproach to
knowledge acquisition provides a framework for
medical experts to build expert systems for
interpreting complexdata sets IF_Awardset al 1993].
For the AIM’94diabetes data set our expert used
TCRDR
to identify features in the data and create his
ownfunctions for use in rules. The flexibility of this
approach further enhances the freedom conferred to
domain experts by RDR, allowing them to build
useful expert systemsin the real world.
The PEIRSresults show that the RDRapproach can
be used with fairly short time course data (five time
points). The results here demonstratethat this can be
extended to more complex time course data. With
more complex time course data it becomes more
difficult to predeterminethe type of data abstraction
required so the TCRDR
system used here allowed the
expert to construct his own function and feature
definitions.
Theexpert building the expert system for the diabetes
data worked on his own with minimal consultation
with the knowledge engineer who originally
packaged the data for him. Most of the consultation
that did take place was in providing the data in the
form the expert wanted, rather than encoding the
experts knowledge. As well, the Interface system
used for the Inference engine (a HyperCardstack in
communication with a separate Inference engine C
program) proved to be a viable method of
communication with the expert. Changes to the
interface were easily madeon the spot as required.
Tasmania, WorldScientific, Singapore 1992, pp. 349354
The crucial features of this developmentare that any
inadequacies in the global functions and features
decided by the expert can be refined in context.
Secondly the expert does not need knowledge
engineering assistance or skill to add rules or
functions in the developing system. Thirdly the
system allows trivial maintenance and on-going
developmentwhile in use. This allows the system to
evolve as more knowledge is accumulated about the
domainand clinical practice evolve.
4. Mansuri,Y., Compton, P., and Sammut,C.: A
comparison of a manual knowledge acquisition
method and an inductive learning method. In: J.
Boose, J. Debenham, B. Gaines and J. Quinlan
(eds.):Australian workshopon knowledgeacquisition
for knowledge based systems. Pokolbin, 1991, pp.
114-132
5. Srinivasan, A., University of NSW& St Vincents
Hospital, Sydney, An expert system for interpreting
chemical pathology data, 1991, Internal Report
In a conventional system changing feature definition
rules causes major
problemsbecause it is very hard to
understand rule interactions (Grossner, Preece et al.
1993). The TCRDR
system solves this problem by
allowing refinement in context. However the
situation mayarise with a particularly poor definition
that the same refinement is being used In many
contexts. This can be solved by a new definition,
howeverthe expert mayprefer to change the original
definition, in which case all the conventional
problems emerge. Westress that this problem has not
yet emerged,but it ts possible. To solve this problem
we have a system under development whereby
changes in definitions are only used by rules added
after the definition has been changed(Menzies1992).
As with RDR,definition changes are validated by
taking into accountcase differences.
6. Edwards,G., Compton,P., Malor, R., Srinivasan, A.
and Lazarus, L. (1993). "PEIRS: a pathologist
maintained expert system for the interpretation of
chemical pathology reports." Pathology : 25:27-34.
7. Clancey, W. (1993). Situated
Action:
Neuropsychological Interpretation Response to Vera
and Simon. Cognitive Science 17(1): 87-116.
8. Compton, P., Kang, B., Preston,
P. and
Mulholland, M. (1993). Knowledge Acquisition
without Analysis, in N. Aussenac, G. Boy, B.
Gaineset al (Eds.), Knowledge Acquisition for
Knowledee Based Systems. Lecture Notes in AI
(723). Berlin, Springe-r Verlag. 278-299.
Finally, this primitive approach of patching errors
seemsinelegant in contrast to the search for elegant
abstractions the characterise someother approaches.
Howeverit is motivated by the "situated cognition"
concern that knowledge is constructed in context,
rather than a Platonic archetype (Comptonand Jansen
1990; Clancey 1993). It has the pleasing
consequence of largely removing the need for a
knowledge engineer and may have wide application
beyond classification systems (Compton,Kanget al.
1993).
9. Compton, P. J. and Jansen, R. (1990).
philosophical basis for knowledge acquisition.
KnowledgeAcquisition 2: 241-257. (Proceedings of
the 3rd European Knowledge Acquisition
for
Knowledge-BasedSystems Workshop, Paris 1989, pp
75-89)
10. Grossner, C., Preece, A., Chander, G.,
Radhakrishnan,T. and Suen, C. (1993). Exploring the
structure of rule based systems. Proceedings of the
American Association of Artificial Inte)ligence,
Washington, MITPress, Cambridge. 704-709
Acknowledgments
This work has been supported by the Australian
Research Council
11. Hauser, T., Kraegen, E. W., Campbell, L. V.,
Compton,P. J. L., Sammut,C. and Chisholm, D. J.
(1992). Assessment of experts’ approach to insulin
therapy and the development of a simulator for
diabetes insulin adjustment. Diabetes Care 15: 221231. (Diabetes Care)
References
1. Compton,P., Edwards, G., Srlnivasan, A., Malor,
R., Preston, P., Kang, B. and LLazarus, L.: Ripple
down rules: turning knowledge acquisition Into
knowledge maintenance. Artificial Intelligence in
Medicine 4,47-59 (1992)
12. Menzies, T. (1992). Maintaining Procedural
Knowledge: Ripple Down Functions.
AI’92.
Proceedingsof the 5th Australian Joint conference on
artificial intelligence, Hobart, -World Scientific,
Singapore. 335-342
2. Compton,P and Jansen, R: A philosophical basis
for knowledgeacquisition. KnowledgeAcquisition 2
, 241-257(1990)
3. Gaines, B. and Compton,P.: Induction of Ripple
DownRules. In: A. Adamsand L. Sterling (eds.):AI
’92. Proceedings of the 5th Australian Joint
Conference on Artificial
Intelligence.
Hobart,
118