Content uploaded by Stephen Beale
Author content
All content in this area was uploaded by Stephen Beale on Jan 01, 2014
Content may be subject to copyright.
Integrating Cognitive Simulation into the
Maryland Virtual Patient
Sergei NIRENBURGa, Marjorie MCSHANEa, Stephen BEALEa,
Bruce JARRELLb and George FANTRYb
aUniversity of Maryland Baltimore County, Baltimore, Maryland
bUniversity of Maryland School of Medicine, Baltimore, Maryland
Abstract. This paper briefly describes four cognitively-related aspects of
modeling a virtual patient: interoception, decision-making, natural language
processing and learning. These phenomena are treated within the Maryland
Virtual Patient simulation and training environment.
Keywords. simulation, cognitive simulation, physiological simulation, medical
training
1. Overview
The Maryland Virtual Patient 1 (MVP) project is developing an agent-oriented
environment for automating certain facets of medical education that includes a network
of human and software agents. The human agents include the user (typically, a trainee)
discharging the duties of an attending physician and, optionally, a human mentor. The
human-like software agents include the virtual patient, lab technicians, specialist
consultants and a mentoring agent. The system also includes an array of non-
humanlike software agents.
At the core of this network is the virtual patient – a knowledge-based model and
simulation of a person suffering from one or more diseases [1-4]. The virtual patient is
a “double agent” in that it models and simulates both the physiological and the
cognitive functionality of a human. Physiologically, it undergoes both normal and
pathological processes in response to internal and external stimuli. Cognitively, it
experiences symptoms, has lifestyle preferences (a model of character traits), has
memory (many of whose details fade with time), and communicates with the human
user about its personal history, symptoms and preferences for treatment.
Users can interview a virtual patient; order lab tests; receive the results of lab tests
from technician agents; receive interpretations of lab tests from consulting physician
agents; posit hypotheses, clinical diagnoses and definitive diagnoses; prescribe
treatments; follow-up after those treatments to judge their efficacy; follow a patient’s
condition over an extended period of time, with the trainee having control over the
speed of simulation (i.e., the clock); and, if desired, receive mentoring from the
automatic mentor.
1 Patent pending.
The virtual patient (VP) simulation is grounded in an ontologically-defined model
of human anatomy and physiology. Instances of virtual patients with particular
diseases and particular physiological traits are generated from ontological knowledge
about human physiology and anatomy by grafting a disease process onto a generic
instance of a human. Disease processes themselves are described as complex events in
the underlying ontology.
We have reported progress on this project at past MMVR meetings. Specifically,
we described our approach to physiological simulation [3] and the way we can
succinctly describe the main features of each disease model to developers, other
experts and the wider community [4]. We continue this year by discussing newly
integrated cognitive capabilities of the virtual patient. The cognitive side of the VP
currently models several aspects of cognitive processing:
o interoception – the perception of physiological phenomena, such as symptoms,
and the interpretation and remembering of such phenomena
o decision making
• deciding when to go see a physician, both initially and during treatment
• deciding whether to seek help in making decisions related to treatment by
asking the user knowledge-seeking questions about a recommended test or
intervention
• deciding whether to agree to a recommended test or intervention
o natural language processing
• language perception and understanding, including both the direct meaning of
physician-user communication in natural language and its intent
• deciding on what specifically to communicate to the user
• actually generating natural language utterances
o learning – receiving new knowledge about the world and the words and phrases
used to describe it and adding them to the ontology and lexicon, respectively
We will briefly touch upon each of the above points with the goal of describing what
the virtual patient can do rather than how it can do it. A sufficient discussion of the
latter would require far more space.
2. Interoception
Interoception is the perception of physiological phenomena. It is a VP feature that has
both physiological and cognitive aspects. The source of interoception is physiological
phenomena, like symptoms of a disease, hunger and sleepiness. Here we focus on
physiology symptoms because our disease models to date have not required the
tracking other kinds of interoception.
The VP experiences current symptoms of its disease and has memories of previous
symptoms, including their severity, so that useful comparisons can be made: e.g.,
“Symptom X has gotten much worse over the past month, I had better go see my
doctor sooner than our next scheduled appointment.”
Memories are stored using an ontologically grounded metalanguage that is
identical to the one used to represent the meaning of language input (cf. below). Of
course, when memories about interoception are stored, there need be no translation into
and from a natural language: the entire process occurs at the level of the metalanguage.
The experiencing of symptoms is individualized for each VP instance through the
use of character traits and physiological features. Our current inventory of character
traits includes trust (trust in the doctor’s advice), suggestibility (how readily the VP
agrees to the doctor’s recommendations) and courage (how willing the VP is to
undergo tests or procedures even if they are risky or have significant side effects). Our
current inventory of physiological traits includes physiological-resistance (e.g., how
well the patient tolerates treatments), pain threshold (how much pain the VP can stand)
and the ability to tolerate symptoms (how intense or frequent symptoms have to be
before the VP consults a doctor). This inventory clearly must be expanded in the future.
When a given VP is created, values for these features are selected and affect the VP’s
reactions in the face of its disease(s). Of course, values for the physiological aspects of
the disease(s) and the VP’s response to interventions, should they be applied at various
times, are also selected for each individual VP.
3. Decision Making
The decision-making behavior of specific instances of virtual patients is parameterized
using a model of personality traits and physical and mental states. It is informed by (a)
the content of the VP’s short-term memory, which is modeled as knowledge invoked
specifically for making the decision at hand, and (b) the content of the VP’s long-term
memory, which is the VP’s recollection of its past states of health, past
communications and decisions, and general world knowledge.
A VP’s decision-making, as described above, is affected by the severity and
duration of its symptoms; its knowledge of tests and procedures; the character traits
trust, suggestibility and courage; and the physiological traits physiological-resistance,
pain-threshold and the ability to tolerate symptoms.
VP reasoning is carried out through modeling the VP’s goals and plans, thus
broadly conforming to the belief-desire-intention (BDI) approach to developing
intelligent agents [5] (see also [6] for related discussion).
When the VP starts to experience symptoms it can either do nothing, go the doctor,
go to the emergency room or self-treat. As concerns seeing a doctor for the first time,
the VP compares its symptom severity with its ability to tolerate symptoms – as well
as some character traits not yet incorporated, such as its desire to be fussed over by a
doctor vs. its dislike of seeing doctors. Later in its treatment the VP also considers the
date of its next scheduled appointment, whether or not its symptoms have spiked, etc.
When the doctor recommends a test or procedure, the VP must compare its
knowledge of the test/procedure with its character traits, like courage and suggestibility.
For example, if it knows nothing about the test/procedure and has little trust in the
doctor, it will ask questions about the properties that interest it, like the pain level and
side effects; by contrast, if it has complete trust in the doctor and a high value for
suggestibility, it will ask no knowledge-seeking questions and, instead, agree to
anything the doctor suggests.
When the patient has received all the knowledge of tests/procedures it feels it
needs, it will decide whether or not to agree to the test/procedure. It can also suggest
other options that it happens to know about, and the doctor can accept or reject such
suggestions.
4. Language Processing
Our approach to treating language communication is unlike most other approaches in
that all language-oriented reasoning is carried out on the basis of formal interpretations
of the meaning of linguistic expressions. Our automatically generated, semantically-
oriented text meaning representations (TMRs) are written using the same ontological
knowledge substrate and the same ontologically grounded metalanguage as are used to
represent physiological processes, interoception and agent goals and plans. In short, all
knowledge and reasoning in our environment employ the same metalanguage, so
whether a VP experiences new symptoms (through interoception) or learns information
about its disease from the user (through language processing), the new information will
be stored the same way in the VP’s memory.
There are several advantages to orienting an agent’s language processing around
TMRs rather than text strings. First, TMRs are unambiguous, since linguistic
ambiguity is resolved as the TMRs are being generated. Second, TMRs reduce to a
single representation many types of linguistic paraphrase, be it lexical (esophagus ~
food pipe), syntactic (I will administer it to you ~ It will be administered to you by me)
or semantic (Does the food get stuck when you swallow? ~ Do you have difficulty
swallowing?) [7]. Third, TMRs facilitate the detection of which aspects of meaning are
central and which are of secondary importance. As regards paraphrase processing, in
addition to having to resolve linguistic paraphrase, the VP must be able to resolve two
other kinds of paraphrase: a) the reformulation of the representation of physiological
events (e.g., symptoms) in “lay” ontological terms that can be understood and
remembered by its cognitive agent and b) the representation of the meaning of verbal
messages in terms compatible with how related content is stored in the cognitive
agent’s memory [8]. Remember, the VP is typically not a medical professional,
meaning that it must have a different ontology and a different lexicon than a physician
would.
5. Learning
We just noted in passing that the VP’s ontology and lexicon do not match those of a
physician. Indeed, the physician’s ontology will contain a vast subtree of medical
information including objects, events and the properties that link them as well as
script-based knowledge [9], which permits the physician to understand the progression
of a disease, how to treat it under various circumstances, etc. The physician will have a
correspondingly large technical and non-technical vocabulary (lexicon) that is linked to
the respective ontological concepts and is used to analyze and generate language in the
medical domain. The patient’s knowledge base, by contrast, will typically include an
impoverished medical subtree in the ontology and a relatively small number of medical
terms in the lexicon – unless, of course, the VP happens to be a physician or even a
specialist in the given domain, which raises a different set of complications for the user.
In conducting an interview with a VP, the user must be able to express himself in
different ways, using paraphrases selected according to the degree of medical
knowledge the VP possesses. However, during appointments the physician will
naturally teach the VP about various aspects of its condition: its name, the names of
related drugs and procedures, the properties of drugs and procedures that the VP asks
about or the user chooses to provide, the medical terms for words that formerly had to
be paraphrased for the VP, and so on. For example, (a) when the user tells the VP the
name of its disease, that disease is added to the VP’s ontological subtree of diseases
and a new lexicon entry is created that maps to this ontological concept; (b) when the
VP learns information about a test or procedure it remembers it and no longer asks
questions about it – unless, of course, the VP forgets, in which case the user will need
offer a reminder. (See related discussions in [10-12].)
6. Discussion
This paper has briefly described four cognitive aspects of VP modeling: interception,
decision-making, natural language processing and learning. Although each of these
presents significant challenges, the work has been facilitated by using the same
ontological substrate for all aspects of modeling.
MVP is a classical AI (artificial intelligence) system in that it strives to model
human perception, reasoning and action capabilities and does so on the basis of
encoded knowledge. It differs from much of classical AI practice in that it includes
people as components in its architecture. If MVP were an expert system in the classical
sense, the system would have been tasked to diagnose and treat patients rather than the
other way around. Indeed, many systems in the medical domain, from Mycin [13] on
up, had this as their main goal. Another difference from classical AI is the centrality of
the descriptive component of the system: the VP’s world is certainly not toy (i.e.,
covering an extremely small domain). Our emphasis is on acquiring knowledge that is
sufficiently deep to support the complex reasoning, simulation and language
processing required by the application. This is in contrast to many recent and current
approaches (notably, in natural language processing) that stress broad coverage of data
in contrast to the utilization of a depth of acquired knowledge.
We believe that the statements often heard nowadays about the demise of AI are ill
conceived. The AI enterprise did not fail. In fact, it has not yet been brought to the test.
This is because the enterprise is much more complex than it was perceived to be even
by many AI practitioners themselves. Despite the recent emphasis in the field on
statistics-oriented methods, they should not be viewed as having superceded classical
AI. In fact, these methods have contributed to the core task of knowledge acquisition
that is a prerequisite to the success of the program of AI. Progress in learning,
knowledge visualization and other ergonomic factors, the ease of access to vast
collections of data on the Web and other developments make the original AI goals
incrementally more attainable. Our work on the MVP corroborates this state of affairs.
We believe that the development of a comprehensive MVP is feasible both
scientifically and logistically.
References
[1] M. McShane, S. Nirenburg, S. Beale, B. Jarrell, G. Fantry, Knowledge-based modeling and simulation
of diseases with highly differentiated clinical manifestations, Proceedings of the 11th Conference on
Artificial Intelligence in Medicine (AIME 07), Amsterdam, The Netherlands, July 7-11, 2007.
[2] M. McShane, G. Fantry, S. Beale, S. Nirenburg, B. Jarrell, Disease interaction in cognitive simulations
for medical training, Proceedings of the MODSIM World Conference, Medical Track, Virginia Beach,
Sept. 11-13, 2007.
[3] B. Jarrell, S. Nirenburg, M. McShane, G. Fantry, S. Beale, D. Mallott, J. Razcek, An interactive,
cognitive simulation of gastroesophageal reflux disease, Stud Health Technol Inform (2007) 194 –199,
IOS Press.
[4] B. Jarrell, S. Nirenburg, M. McShane, G. Fantry, S. Beale, Revealing the conceptual substrate of
biomedical cognitive models to the wider community, Stud Health Technol Inform (2008), 281 – 286,
IOS Press.
[5] A.S. Rao, M.P. Georgeff, Modeling rational agents within a BDI-architecture, Proceedings of the 2nd
International Conference on Principles of Knowledge Representation and Reasoning, Morgan
Kaufmann Publishers, Inc., San Mateo, CA, USA, 1991, 473 – 484.
[6] S. Nirenburg, M. McShane, S. Beale, A simulated physiological/cognitive “double agent”, Proceedings
of Naturally Inspired Cognitive Architectures, AAAI 2008 Fall Symposium, Washington, D.C., Nov. 7-
9, 2008 (forthcoming).
[7] M. McShane, S. Nirenburg, S. Beale, Resolving paraphrases to support modeling language perception
in an intelligent agent, Proceedings of the Symposium on Semantics in Systems for Text Processing
(STEP 2008), Venice, Italy, 2008.
[8] M. McShane, S. Nirenburg, S. Beale, Two kinds of paraphrase in modeling embodied cognitive agents,
Biologically Inspired Cognitive Architectures, AAAI 2008 Fall Symposium, Washington, D.C., Nov. 7-
9, 2008 (forthcoming).
[9] R. Schank, R. Abelson, Scripts, Plans, Goals and Understanding, Hillsdale, NJ, Erlbaum, 1977.
[10] S. Nirenburg, T. Oates, Learning by reading by learning to read, Proceedings of the International
Conference on Semantic Computing (ICSC-07), Irvine, CA, 2007.
[11] P.Z. Yeh, B. Porter, K. Barker, Using transformations to improve semantic matching, Proceedings of
the 2nd International Conference on Knowledge Capture, Sanibel Island, FL, USA, 2003, 180–189.
[12] S. N irenburg, M. McShane, S. Beale, B. Jarrell, Adaptivity in a multi-agent clinical simulation system,
Proceedings of AKRR’08, International and Interdisciplinary Conference on Adaptive Knowledge
Representation and Reasoning, Porvoo, Finland, Sept. 17 – 19, 2008.
[13] B. Buchanan, E. Shortliffe, Rule-Based Expert Systems: The Mycin Experiments of the Stanford
Heuristic Programming Project, Addison-Wesley, 1984.