Three Compatible Mechanisms for Representing Medical Context Implicitly

Three Compatible Mechanisms for Representing Medical Context
Implicitly
Michel DOJAT1, François PACHET2
1INSERM Unité 296, Faculté de médecine, 94010 Créteil, France.
2LAFORIA-IBP, Université Paris 6, Boîte 169, 4, Place Jussieu, 75252 Paris,
France.
e-mails: dojat/pachet@laforia.ibp.fr
Abstract
An important aspect of context in medical reasoning is the notion of
"variation" of a chunk of knowledge according to various contingencies, such as
course of patient's disease, response to therapies, or team specificity. Our
position is to represent these variations implicitly, by proposing mechanisms to
factor knowledge and to refine it. We propose three mutually compatible
mechanisms that effectively contribute to represent slight variations of
knowledge in a representation framework integrating object-oriented
programming and rule-based programming. We illustrate them with examples
extracted from various knowledge bases for the management of mechanical
ventilation.
Key words: Context, Knowledge Representation, Object-Oriented
Programming, Rule-Based Programming.
1. Introduction
In medical reasoning, the notion of context is central for interpreting
physiological parameters and is relevant at all levels of expertise. [Brézillon 93]
showed that, within the community of researchers working on context issues,
two main tendencies emerge:
1) Context as a "situation"
Context considered as something (ill-defined) designating the situation in
which an interaction occurs, i.e. the place, the task, the participants, etc.
In the case of medical knowledge, context as situation encompasses several
dimensions:
- context for the interpretation of data. Context reflects patient's history, type of
disease or response to therapies.
- context to account for various temporal phases in the course of the patient's
disease. Phases may correspond to various tasks occurring over time. In each
phase, specific reasoning is envisaged.
- context to account for the specificity of the clinical team. Although clinical
practice is largely based on consensual medical knowledge, each clinical team,
specialized in a particular medical field, accumulates specific experience. This
difference with commonly accepted knowledge reflects habits, cultural
contexts, specific skills and experiences.
1

---

2) Context as a "set of beliefs"
Context considered as a restricted set of beliefs or logical assumptions that are
necessary to interpret correctly available information. In the medical field, it is
often important to explicitly represent the underlying assumptions of the
medical doctor regarding the patient's status, such as the diagnosis established,
the theoretically predicted evolution of the patient's state or the expected
evolution in function of the current treatment.
In medical expertise, knowledge to be used in several contexts bear a lot of
similarities. As we will see in the examples, a vital aspect of context is the
notion of variation of a chunk of knowledge according to various contingencies,
such as patient's response to therapy, course of patient's disease or team
specificity. Representing similarities between different contexts raises complex
issues. For instance, [Guha 91] proposes a dedicated mechanism (the so-called
lifting rules) to establish some kind of communication between different but
related contexts.
Moreover, context is multiform and the mechanisms proposed to account
for it in the representation of knowledge, are various and often difficult to
conciliate. For example, the context mechanism proposed by [Guha 91] is hard
to transpose in a non-logical representation framework. Similarly, the
mechanism proposed by [Abu-Hakima 93] requires the explicit representation
of goals to be used by the inference process.
Based on these two remarks, our position regarding context is 1) to focus on
mechanisms, rather than on explicit representations, and 2) focus on similarities
rather than on dissimilarities, and 3) ensure some kind of compatibility of
various manifestations of context-dependency. We propose three mutually
compatible mechanisms for factoring medical knowledge, thereby providing a
powerful implicit representation of context. These three mechanisms are: class
inheritance, natural typing and rule base inheritance. They are applicable in
representation frameworks integrating rule-based programming with object-
oriented programming.
Our research is being conducted in collaboration with a team specialized in
intensive care medicine at Henri Mondor Hospital (Créteil, France). We
developed a knowledge-based system for ventilator management, called
NéoGanesh. NéoGanesh adjusts in closed-loop the mechanical assistance,
depending on the patient's physiological needs, and indicates when the patient
can breathe without external assistance (weaning). The results [Dojat et al. 92],
[Dojat et al. 94] demonstrated that NéoGanesh improves the quality of the
patient's ventilation, and therefore validates our approach. To illustrate the
interest of our proposals, we will use examples extracted from NéoGanesh and
from expertise in the same area found in the literature.
2. Our representation framework: Alliance of Objects and Rules
Object-oriented languages propose to represent knowledge by simulation.
Software objets are constructed to represent a monitor, a patient, or a clinician,
as well as therapies, diagnosis, expectations and so forth. Operations on these
2

---

objects represent domain operations such as monitoring respiratory rate,
changing respiratory rate and administrating a new therapy, and are
represented as methods associated to corresponding classes.
The need for combining object structures and rule-based programming has
been widely recognized. The fact base of a rule-based program is a model of
the concrete situation that is currently being processed. To bring some
semantic structure to facts, one naturally tends to see them as properties of
objects that build up a universe simulating the concrete world. Individual facts
are no longer represented as such, their logical value is ascertained by querying
objects in the model. The fact base is thus dissolved into an object-oriented
model of the world. The equation "fact base = model of the world" links the
two techniques.
Starting with a standard object-oriented language (Smalltalk-80), and
extending it with an embedded rule-based layer, we built the NéOpus system
[Pachet 95]. Turning to classical object-oriented style (the so-called message
passing) causes some trouble to the knowledge representation specialist (see,
e.g. [Nebel 90]). We refer to [Pachet 94] for a discussion on the technical
problems posed by this integration. However, we feel that the benefit gained
from potentially applying rules to the whole universe of object-oriented models
created by object-oriented programmers does warrant these discrepancies from
the generally accepted principles.
3. Three Mutually Compatible Mechanisms to Account for
Context
We propose three mutually compatible mechanisms that address three
different levels of granularity for factoring medical knowledge representing
context in our object-oriented setting 1) class inheritance, 2) natural typing, and
3) rule base inheritance.
3.1. Class Inheritance to Account for Context
The organizing principle of class inheritance allows to describe medical
information and create information structures comprising concepts which are
statically related through common property characteristics. Although class
inheritance can hardly be compared with true classification mechanisms as
found, e.g. in description logics [Patel-Schneider 90], it nevertheless allows us
to represent taxonomies of physical or conceptual entities and
generalization/specialization relation (e.g. class Intensivist inherits from
class Clinician which in turn inherits from MedicalDoctor). Here is an
example that shows how class inheritance can be used effectively to account for
context.
Availability of data as Context
Well-structured problems of diagnosis are solved by the method of heuristic
classification. This method requires three steps: 1) abstraction from input data,
2) association with a taxonomy of possible solutions and 3) refinement of the
collection of potential solutions [Clancey 85]. Embedding classification
3

---

mechanisms in object-oriented languages is not, however a trivial task [Yelland
92]. For our applications, we propose to use class inheritance to represent the
taxonomy of possible solutions. Each class holds a set of constraints to which
patient's data is matched for classification. For instance, to diagnose the current
respiratory state of a patient, clinicians currently use three main parameters:
respiratory rate (RR), volume inspired at each breathing cycle (Vt), and
pressure of carbon dioxide at the end of the expiration (PetCO2).
Consequently, the class Normal representing a normal ventilation, holds the
following set of constraints (expressed as a single Boolean expression):
Constraint for class "Tachypnea"
VT > 250 AND
(RR < 35 AND RR > 28) AND
PetCO2 < 55
The classification process consists in matching a given patient ventilation
state (represented by an instance of class MeasuredVentilation) against the
solution classes. For instance, the solution classes that require parameters VT,
RR, and PetCO2 are: Normal, Tachypnea, Bradypnea, Insufficient, and so
forth. The definition attached to each element of the taxonomy of solutions, as
well as the thresholds used for interpreting physiological data, may vary
depending to the context: the patient's characteristics (sex, age, pathology, etc.),
the clinical evolution, or the medical doctor. Moreover, the refinement of the
potential solutions may depend on the availability of certain data. For instance,
to characterize the patient's respiratory state more precisely, we can add a
parameter, called Occlusion Pressure or P0.1, which measure the patient's effort
needed to trigger the mechanical assistance. To represent this new piece of
information (and hence a new context of interpretation of data), we add new
solution classes to the taxonomy, with constraints that explicitly refer to this
new data. We add the solution class FineTachypnea to our taxonomy by
defining a subclass of Tachypnea. Similarly, we will subclass Normal with
class FineNormal, holding a constraint with an additional condition on P0.1,
and so forth.
Instead of systematically matching against all the classes of the taxonomy,
we organize solution classes in several collections, each one representing a
particular combination of available parameters. These collections of classes are
represented as methods in the class that represents the current clinician
(Intensivist). standardSolutionsList represents the list of solution
classes when RR, Vt and PetCO2 are available, fineSolutionsList is used if
P0.1 is available and so forth.
3.2. Natural Typing to Account for Context
The preceding example shows how inheritance may be used to factor
information. Encapsulation (or information hiding) supplies a useful barrier
among several levels of abstraction. Now class inheritance and encapsulation
may be used in conjunction with rules, yielding another dimension of context
representation. This is realized in NéOpus by so-called natural typing of rule
variables.
4

---

3.2.1. The Mechanism
The idea behind natural typing is to allow the pattern-matcher to consider
direct instances as well as instances of subclasses to be matched by rule
variables for a given rule. The interpretation of a rule is therefore dynamic, since
1) the condition and action parts of the rules are entirely expressed in terms of
messages sent to the matched objects and 2) the messages are redefined in
subclasses. In other words, the rules are context-dependent, where the context is
represented by the set of objects that match the rule.
Class1
Class2
Class3
mess1
mess2:
mess3
• mess1
•
•
• mess2:
• mess3
•
•
•
•
• mess2:
•
•
•
•
•
mess1
r1
| Class1 v1. Class2 v2. Class3 v3 |
If v1 mess1.
v2 mess2: v3.
Then
v3 mess3.
Figure 1. A rule with the corresponding class hierarchies. The rule should be read as follows:
For any v1, instance of class Class1, v2 instance of Class2 and v3 instance of Class3, if "v1
mess1" and "v2 mess: v3" evaluate to true, then evaluate "v3 mess3". The rule uses three
messages (mess1 through m3) which are defined and redefined in the classes of the
corresponding hierarchies.
More precisely, let r be a rule that declares n variables vi (i=1, n). Let Ci be
the class declared for vi. Now, if each Ci has ki concrete subclasses (Cf. Figure
1), the total number of possibly different interpretations of r is Π ki (i = 1, n).
In practice, if r has 3 variables, and each declared class has 5 concrete
subclasses, then r has 53 = 125 possibly different interpretations !
3.2.2. Reference Values as Context
[East et al. 94] develop computerized protocols to assist clinicians with the
difficult task of ventilating mechanically patients with Acute Respiratory Distress
Syndrome (ARDS). They identify the arterial oxygenation of the patient by
means of PaO2 measurement to define the best therapy to provide. If the
clinician diagnoses a Barotrauma, he tolerates a lower arterial pressure of
oxygen, and tries to reduce alveolar pressures. We represent references of PaO2
by classes (Cf. Figure 2, left), and their variations by subclasses. In this scheme,
the reference used for PaO2 in the case of Barotrauma is therefore represented
by a class PaO2Low, subclass of PaO2Reference, which redefines method
"low" with the appropriate value.
The chosen strategy, function of the diagnosis performed, determines the
appropriate subclass of PaO2Reference that should be used for classifying
PaO2. The actual arterial oxygenation of the patient is itself represented by an
instance of MeasuredVentilation. The rule (Cf. Figure 2, right), is then used
to classify arterial oxygenation by comparing it with the chosen PaO2 reference.
5

---

Thanks to natural typing, the condition part of the rule is interpreted
differently depending on the class of the reference of PaO2.
PulmonaryReferences
defines:
acceptableOxygenation
low, moderate, high
| M e a s u r e d V e n t i l a t i o n v e n t .
Pa02Reference reference|
PaO Reference
defines:
2
veryLow, veryHigh
vent paO2 > reference low.
vent paO2 < reference moderate.
PaO Low
2
PaO High
2
then
redefines: low, moderate
redefines: high, moderate
Oxygenation new acceptable.
Figure 2: Left: gradual introduction of modification of thresholds for arterial pressure of oxygen
classification. Right: an example of rule to classify arterial oxygenation. Methods low and
moderate, which return corresponding threshold values, are redefined in subclasses of
PaO2Reference.
3.2.3. Management of Time
In most medical knowledge-based systems, the course of the patient's
disease is divided into several temporal phases. These phases may correspond
to steps in the treatment, e.g. in cholesterol treatment [Rucker et al. 90], or
respiratory assistance [Fagan 80], or various states of the patient's evolution,
such as hemodynamic states ([Lau and Vincent 93], [Cohn et al. 90]). In each
phase, specific rules are applied.
Recently, [Dojat and Sayettat 94] proposed a model for temporal reasoning
representation in real-time systems. This model is based on temporal
abstractions that allow observations to be interpreted incrementally as they are
acquired. Two mechanisms are used: aggregation of similar situations and
forgetting non-relevant, redundant or out of date information. Activation of
these two mechanisms are context-dependent.
We implement this model using our combination of class inheritance and
rules. This is performed by 1) reifying phases, and building a hierarchy of
classes representing various significant phases of the course of the patient's
disease and 2) introducing an explicit management of time that account for
state change and evolution.
TemporalObject
State
RespiratoryState
ExpectedState
Maintain
Restoration
PreWeanable
Weanable
Figure 3: Excerpt of the hierarchy of temporal objects in NéoGanesh.
6

---

Temporal objects are time-stamped entities, used to develop a temporal
discourse about the patient's ventilation evolution. The figure 3 shows parts of
the hierarchy of temporal objects used in NéoGanesh. RespiratoryState
represents the ventilation status of the patient. ExpectedState represents the
future patient's respiratory state, as expected by the doctor. One of the goal of
the medical reasoning is to appreciate the stability of the ventilation over
periods of time, the duration of instabilities and the persistence of inadequate
therapies. Thus, the clinician has to perceive significant changes between
successive observed states or successive expected states. This perception of
changes is, of course, context-dependent: for instance, the notions of similarity
and dissimilarity between states depends on the class of the states compared.
This is, once again, represented by methods attached to the corresponding
subclasses of TemporalObject. Figure 4 shows a rule taken from the rule base
that handles perception of change.
partialContinuity
|State s1 s2 s3. Duration d1|
s3 persistent.
s1 similarTo: s3.
s2 dissimilarTo: s3.
s2 between: s1 and: s3.
s2 duration <= d1.
then
s1 validDuring: s3 duration.
Figure 4: A rule that aggregates two disjoint states (s1 and s3), separated by an instability (s2)
considered short enough to be discarded.
NéoGanesh includes numerous other examples of the use of natural
typing, including more sophisticated management of thresholds and
representation of control ([Dojat and Pachet 92]).
3.3. Rule Base Inheritance to Account for Context
We introduce a third level of context-dependency: rule base inheritance
(RBI). RBI allows to specify gradually the context of application of rules by
matching objects that are relevant for this context.
3.3.1. The Mechanism
In our representation framework (NéOpus), rules are grouped in rule bases
that are represented as abstract classes. The idea of transposing the class
inheritance mechanism to rule bases is therefore natural. We introduced in
[Pachet and Perrot 94] a scheme to transpose the intuition of class inheritance to
rule bases. In this scheme, each rule base may be defined as a sub-base of an
existing rule base, thereby inheriting all its rules. An overriding mechanism,
based on rule names, allows a rule base to redefine a inherited rule into a more
specific rule. This scheme is based on a static propagation of rule compilation
to sub-bases, together with a particular control strategy (RBI strategy) that
ensures that, in case of conflict, rules defined in the lowest sub-base will be
selected.
There are indeed several advantages in providing rule bases with an
inheritance mechanism. It provides a high level scheme for organizing rules, it
allows to factor out common rules, and simplifies control strategy specification.
7

---

We will show here how RBI may be used to provide explicit contextual
dependency of medical knowledge.
3.3.2. RBI for Context Dependency in Temporal Reasoning
The Figure 5 shows the inheritance tree for rule bases, which gradually
introduce context-dependency in temporal management, at the rule base (or
task) level.
Rule base TemporalManagement contains rules that define the perception
of continuity, discontinuity, partial continuity and so forth, in a use-neutral
manner. They match general temporal objects (State).
Rule base VentilationEvolution introduces a refinement of this general-
purpose temporal reasoning to adapt it to the ventilation management.
VentilationEvolution inherits all the rules defined in its super-base,
redefines some of them (such as partialContinuity), and defines specific
rules (such as extendedAggregation). Rules defined or redefined in
VentilationEvolution talk about RespiratoryState (instead of the
general State class, see Figure 6).
NéOpusRuleBase
TemporalManagement
continuity
partialContinuity ...
VentilationEvolution
partialContinuity
extendedAggregation ...
WeaningEvolution
extendedAggregation
successfulWeaning ...
Figure 5: An inheritance tree of rule bases.
partialContinuity (redefined in VentilationEvolution)
|RespiratoryState s1 s2 s3|
s3 persistent.
s1 similarTo: s3.
s2 dissimilarTo: s3.
s2 between: s1 and: s3.
s2 durationInExpertise <= 1.
s3 durationInExpertise > 1.
then
s1 validDuring: s3 duration.
Figure 6: Rule partialContinuity redefined in a sub-base of TemporalManagement. The
rule matches more specialized objects (RespiratoryState instead of State), modifies and
adds a condition. The rule states that 2 disjoint respiratory states, separated by an instability
that lasts less than 1 expertise cycle, can be aggregated.
Finally, WeaningEvolution is concerned by situations when the patient is on
the verge of breathing without assistance. ExtendedAggregation rule is
modified to tolerate moderate ventilation instabilities. SuccessfulWeaning
8

---

indicates to the user that the patient can be extubated. At this level, rules match
only instances of class Weanable.
4. Discussion
We have introduced three mutually compatible mechanisms that take
context into account. We illustrated the mechanisms by instanciation of
context-dependency for availability of data (class inheritance), representing
reference values (natural typing in forward-chaining rules), and representing
different perception to changes (rule base inheritance). Object-orientation casts a
new light on the distinction between the two notions of context (situation
versus set of beliefs). Because our representation model is uniform (everything
is represented by an object), the notion of a "logical fact" disappears, making it
difficult to represent "sets of beliefs". However, we think that this distinction is
still relevant, and propose to take it into account by distinguish between two
categories of domain objects:
- perceived objects (symptoms, devices, patient's status and history), that
represent the world as it is "perceived" by the clinician expert. These objects
represent concrete real-world objects as well as objects pre-existing to the
reasoning process.
- conceived objects (diagnosis, models of ventilation and expectations) that
represent the world as it is "conceived" by the clinician expert, and created or
worked out by his reasoning process.
In this setting, the distinction situation/belief is revisited: situation is
represented by perceived objects, and beliefs by conceived objects.
Few systems propose precise descriptions of the implementation level and
how they practically achieve some forms of context representation.
In this
paper we identified three mechanisms of rule-based object-oriented
programming: class inheritance, natural typing and rule base inheritance and
showed how they allow to implement various dimensions of context-
dependency in medical knowledge bases. This reinforces the potential interest
of the object-oriented paradigm and production rules association. We are
convinced that the development of large open health care systems might
strongly benefit from this approach.
5. References
[Abu-Hakima 93] Abu-Hakima S., The Use of Context in Diagnostic Systems, in
Proceedings IJCAI'93 Workhop on Using Knowledge In its Context, Chambéry,
France, Université Paris 6, pp. 13-19, 1993.
[Brézillon 93] Brézillon P., Proceedings of the IJCAI 93 Workshop on Using
Knowledge in its Context, Technical Report 93/13, LAFORIA-IBP, Université
Paris 6, 1993.
[Clancey 85] Clancey W.J., Heuristic classification, Artificial Intelligence, vol. 27,
pp. 289-350, 1985.
[Cohn et al. 90] Cohn A.I., Rosenbaum S., Factor M. and Miller P.L.,
DYNASCENE: An approach to computer-based intelligent cardiovascular monitoring
9

---

using sequential clinical "scenes", Methods of Information in Medicine, vol. 29,
pp. 122-131, 1990.
[Dojat et al. 92] Dojat M., Brochard L., Lemaire F. and Harf A., A knowledge-
based system for assisted ventilation of patients in intensive care, International
Journal of Clinical Monitoring and Computing, vol. 9, pp. 239-250, 1992.
[Dojat et al. 94] Dojat M., Harf A., Touchard D., Laforest M., Lemaire F. and
Brochard L., Clinical evaluation of a knowledge-based system providing ventilatory
management and decision for extubation during weaning from mechanical ventilation,
American Review of Respiratory Disease, vol. (submitted), 1994.
[Dojat and Pachet 92] Dojat M. and Pachet F., NéoGanesh: an extendable
knowledge-based system for the control of mechanical ventilation, in Proceedings 14th
annual international conference of IEEE-EMBS, Paris, pp. 920-921, 1992.
[Dojat and Sayettat 94] Dojat M. and Sayettat C., A realistic model for temporal
reasoning in real-time patient monitoring, Applied Artificial Intelligence, (to
appear), 1994.
[East et al. 94] East T., Morris A.H., Wallace C. J., Clemmer T. P., Weaver L.
K. and Orme J. F., Computerized protocols for the management of arterial
oxygenation in patients with ARDS on assist control mechanical ventilation,
American Review of Respiratory Disease, (to appear), 1994.
[Fagan 80] Fagan L.M., Representing time dependant relations in a medical settings,
Ph.D. thesis, University of Stanford (Ca), 1980.
[Guha 91] Guha R.V., Contexts: A formalization and some applications, Technical
Report ACT-CYC-423-91, MCC, 1991.
[Lau and Vincent 1993] Lau F. and Vincent D. D., Formalized decision support
for cardiovascular intensive care, Computers and Biomedical Research, vol. 26, pp.
294-309, 1993.
[Nebel 90] Nebel B., Reasoning and Revision in Hybrid Representation Systems.
Lecture Notes in Artificial Intelligence, Berlin: Springer-Verlag, 1990.
[Pachet 94] Pachet F., Report on the OOPSLA'94 Workshop on EOOPS, in
Proceedings Addendum to the OOPSLA'94 proceedings, Portland, Oregon,
ACM Sigplan notice, 1994.
[Pachet 95] Pachet F, On the embeddability of production systems in object-oriented
languages, Journal of Object-Oriented Programming, (to appear), 1995.
[Pachet and Perrot 94] Pachet F. and Perrot J.-F., Rule Firing with Metarules, in
Proceedings SEKE'94, Jurmala, Latvia, Knowledge Systems Institute, pp. 322-
329, 1994.
[Patel-Schneider 90] Patel-Schneider P. F., Practical, Object-Based Knowledge
Representation for Knowledge-Based Systems, Information Systems, vol. 15, pp. 9-
19, 1990.
[Rucker et al. 90] Rucker D.W., Maron D.J. and Shortliffe E.H., Temporal
representation of clinical algorithms using expert system and database tools,
Computers and biomedical research, vol. 23, pp. 222-239, 1990.
[Yelland 92] Yelland Ph. M., Experimental Classification Facilities for Smalltalk, in
Proceedings OOPSLA '92, Vancouver, Canada, pp. 235-246, 1992.
10

---