Upholding the Maxim of Relevance 
during Patient-Centered Activities 
Abigail S. Gertner and Bonnie L. Webber John R. Clarke 
Computer ~ Information Science 
University of Pennsylvania 
Philadelphia PA 19104-6389 
agertner©linc, cis. upenn, edu 
bonnie@central, cis. upenn, edu 
Abstract 
This paper addresses "kinds and focuses of 
relevance" that a language-generating clin- 
ical decision-support system should adhere 
to during activities in which a health care 
provider's attention is on his or her pa- 
tient and not on a computer screen. Dur- 
ing such "patient-centered" activities, ut- 
terances generated by a computer system 
intrude on patient management. They 
must be thus seen by HCPs as having 
immediate clinical relevance, or, like the 
continual ringing of ICU monitors, they 
will be ignored. This paper describes how 
plan recognition and plan evaluation can 
be used to achieve clinical relevance. The 
work is being done in the context of the 
TraumAID project, whose overall goal is to 
improve the delivery of quality trauma care 
during the initial definitive phase of patient 
management. Given an early pilot study 
that showed that physicians using Traum- 
AID disliked the continuous presentation of 
its entire management plan, we decided to 
explore how TraumAID could restrict com- 
mentary to only those situations in which 
a comment could make a clinically signifi- 
cant difference to patient management. We 
took advantage of the fact that actions that 
involve resources that need to be brought 
to the trauma bay or that can only be 
done elsewhere must be ordered. Since or- 
ders can be rescinded, comments pointing 
out problems with an order can potentially 
make a clinically significant difference to 
patient management. The contributions of 
this paper are (1) pointing out additional 
*This work has been supported in part by the 
Army Research Organization under grant DAAL03-89- 
C0031PRI, the National Library of Medicine under grant 
R01 LM05217-01 and the Agency for Health Care Policy 
and Research under grant RO1 HS06740. The authors 
would fike to thank Mark Steedman and Jonathan Kaye 
for their helpful comments on earlier drafts of this paper. 
Department of Surgery 
Medical College of Pennsylvania 
3300 Henry Avenue 
Philadelphia PA 19129 
j clarke@gradient, cis. upenn, edu * 
constraints on language generation raised 
by the desire to convey information to lis- 
teners attending to something other than 
an computer terminal, and (2) pointing out 
some features of plan inference and evalu- 
ation raised by multiple goal planning in a 
complex domain. 
1 Introduction 
Ordinary use of Natural Language adheres to what 
Grice has called the "cooperative principle of conver- 
sation" (Grice, 1975). Four categories of "maxims" 
elaborate this principle. Under the category of Re- 
lation, Grice places the single maxim "Be relevant", 
noting that this terse formulation conceals a num- 
ber of unsolved problems, including that of "what 
different kinds and focuses of relevance there may 
be". 
This paper addresses "kinds and focuses of rele- 
vance" that a language-generating clinical decision- 
support system should adhere to during patient- 
centered activities - that is, activities in which a 
health care provider's (HCP) attention is on his 
or her patient and not on a computer screen. 1 
Such activities include surgery, childbirth, and emer- 
gency medical care (including the responsibilities of 
Emergency Medical Technicians). During patient- 
centered activities, utterances generated by a com- 
puter system intrude on patient management. They 
must be seen by HCPs as having immediate clinical 
relevance, or, like the continual ringing of ICU mon- 
itors, they will be ignored. Utterances that would 
not make a clinically significant difference to patient 
management must be avoided. This paper describes 
how plan recognition and plan evaluation can be 
used to achieve this end. 
This work is being done in the context of the 
TraumAID project (Clarke et al., 1993; Clarke et 
al., 1994; Rymon et al., 1993; Webber et al., 1992), 
a In this sense, our emphasis differs from Sperber and 
Wilson (Sperber and Wilson, 1986), who are concerned 
with the problem of inferences that can be made under 
the assumption of relevance. 
125 
whose overall goal is to improve the delivery of qual- 
ity trauma care during the initial definitive phase of 
patient management. Initial definitive management 
scopes a complex range of diagnostic and therapeu- 
tic procedures. Resolving often conflicting demands 
for managing multiple injuries requires global rea- 
soning that can exceed the restricted locality of rule- 
based systems. The modular architecture of the cur- 
rent system, TraumAID 2.0, consists of two comple- 
mentary processes: (1) a rule-based reasoner able to 
draw diagnostic conclusions and identify what goals 
each of them implies; and (2) a planner that takes 
the set of currently relevant goals and constructs a 
management plan - a partially-ordered sequence of 
actions - that is most appropriate at that point in 
time. 
TraumAID 2.0 has been retrospectively validated 
against actual trauma management plans for 97 non- 
pregnant adult patients presenting consecutively to 
the Medical College of Pennsylvania (MCP) Trauma 
Center with penetrating injuries to the chest and/or 
abdomen. A blinded comparison of management 
plans was carried out by a panel of three experienced 
trauma surgeons. They preferred TraumAID 2.0% 
plans to actual management to a statistically signif- 
icant extent (Clarke et al., 1993). 
A decision support system will not be clinically 
viable without solving its interface problems. For 
us, this means (1) getting TraumAID the informa- 
tion it needs for its reasoning and planning, and (2) 
getting the results of TraumAID's reasoning to the 
members of the trauma team (in particular, to the 
chief surgical resident) during the course of care. 
To get information into TraumAID, we have de- 
veloped an electronic version of a standard trauma 
flow sheet, so that information can be entered by the 
scribe nurse during the trauma resuscitation. 2 The 
electronic trauma flow sheet is implemented in Hy- 
perCard, supported by a FoxPro database. Relevant 
information entered by the scribe nurse is automat- 
ically passed to TraumAID 2.0. 
The focus of this paper is the problem of delivering 
the results of TraumAID's deliberations, given an 
early pilot study that showed that physicians using 
TraumAID disliked the continuous presentation of 
its entire management plan. (Currently, this is dis- 
played on the nurse's monitor in a separate window.) 
The reasons for physician dissatisfaction appear to 
be that (1) much of the time, TraumAID's recom- 
2One member of the trauma team is a nurse who func- 
tions as a scribe, documenting all the findings, tests, and 
treatments in chronological order for the record. The 
multiple page trauma flow sheet has designated areas for 
specific information, such as demographics, mechanism 
of injury, physician response times, trauma score, Glas- 
gow coma score, vital signs, location of wounds, results 
of primary assessment, intravenous therapy, diagnostic 
and therapeutic procedures, medications given, fluid in- 
take and output, and disposition. 
mendation coincided with the physicians' own plans, 
and (2) having the entire plan presented made it dif- 
ficult for the physicians to determine what, if any- 
thing, they should focus on. To put it another way, 
presenting the entire plan violates Grice's Maxim of 
Relevance. 
We decided to explore how TraumAID could re- 
strict commentary to only those situations in which 
a comment would be clinically relevant in terms of 
its potential for making a clinically significant differ- 
ence to patient management. We took advantage of 
the fact that actions that involve resources that need 
to be brought to the trauma bay or that can only 
be done elsewhere must be ordered. Since orders can 
be rescinded, comments pointing out problems with 
an order can potentially make a clinically significant 
difference to patient management. Our prototype 
interface for TraumAID produces commentary when 
an order differs significantly from what TraumAID 
would recommend. It also generates commentary 
when the physician fails to order something that 
TraumAID thinks should be done within the imme- 
diate time frame. The interface, TraumaTIQ (Gert- 
net, 1994; Gertner, 1994b), uses plan inference and 
plan evaluation to recognize both such errors of com- 
mission and errors of omission. In this paper, we de- 
scribe TraumAID's knowledge of management goals 
and actions that are used in plan inference and eval- 
uation (Section 2), the plan recognition and evalu- 
ation strategies that allow TraumaTIQ to recognize 
clinically significant differences in patient manage- 
ment(Section 3), and the additional problems for 
plan recognition, plan evaluation and critique gener- 
ation caused by a planner's ability to optimize plans 
through what Pollack has called "overloading inten- 
tions" (Pollack, 1991) (Section 4). We conclude with 
mention of our planned prospective evaluations of 
TraumAID and TraumaTIQ and a summary of our 
main points (Section 5). 
2 The Representational Framework 
In TraumAID, a plan is a set of procedures, each of 
which comprises one or more actions. Procedures 
are used to address diagnostic and therapeutic man- 
agement goals motivated by what is currently known 
about the patient. A particular action can partici- 
pate in more than one procedure, and thus can be 
used in addressing more than one goal. As noted 
earlier, TraumAID uses a rule-based reasoner to de- 
rive a set of relevant goals from the currently avail- 
able knowledge about the patient and a planner to 
choose appropriate procedures to address those goals 
and schedules the actions they involve. 
TraumAID's reasoner controls information acqui- 
sition using a conservative, staged strategy for di- 
agnosis and treatment (Rymon, 1993): expensive, 
definitive tests are not included in a plan until they 
are justified by less costly tests or observations, and 
126 
G1 (32 G1 G2 
A1 A2 A3 A4 A5 A1 A2 A3 
Shared Action Shared Procedure 
Figure h Three possible multiple 
GI q 
A1 A2 A3 
G2 
A4 A5 A6 
Independent 
goal-procedure-action configurations 
goals 
procedures 
actions 
definitive treatment is not recommended without the 
results of sufficient evidence from diagnostic tests. 3 
These strategies are reflected in the knowledge base 
by the occurrence of related management goals, such 
as a goal to diagnose hematuria (blood in the urine), 
which if present, triggers a goal to diagnose bladder 
injury, which in turn can lead to a goal to treat blad- 
der injury. Goals that do not participate together in 
a coherent strategy may still be connected by test 
results. For example, the goal of finding a bullet in 
the mediastinum leads to doing a lateral chest x-ray. 
While this might also reveal a fractured sternum, 
there is no strategic relationship between the goal of 
finding a bullet in the mediastinum and the goal of 
treating a fractured sternum. 
Once a set of relevant goals has been determined, 
the planner's choice of how address each goal is 
based on both local and global considerations. Goals 
and procedures are linked via a goal-procedure map- 
ping which lists alternative procedures for address- 
ing each goal. The procedures in a mapping are 
listed in order of preference so that, all else being 
equal, the first procedure will be chosen to address 
the goal. Less preferred procedures may be selected 
if they would result in a globally more optimal (less 
costly) plan. Figure 1 shows three possible configu- 
rations for a portion of a plan addressing two goals. 
(Section 4 discusses the consequences for plan recog- 
nition, evaluation and critiquing, of the ability to 
create plans in which actions and/or procedures are 
shared between goals.) 
Actions are scheduled according to both logistical 
and clinical considerations. Logistical considerations 
mean that patients are only moved in one direction 
through the Trauma Center - from the emergency 
center, optionally to the radiology suite, then op- 
tionally to the operating suite, and finally to the 
3The one case where this is not true is when a patient 
comes in near death, with catastrophic chest wounds. 
Surgery is recommended immediately, without attempt- 
ing to diagnose what specific injuries may have been 
sustained. 
127 
trauma unit. Since actions may have constraints on 
where they can be performed, they are scheduled so 
as to avoid transferring the patient back to a place 
he has already been. 
Clnical considerations have to do with the ur- 
gency and priority of each action, which it inher- 
its from the goals it is being used to address. The 
urgency can be either calastrophic, unstable, or sta- 
ble, representing the patient's condition and thus the 
amount of time available in which to address that 
goal. Catastrophic goals must be addressed imme- 
diately. Unstable goals must be addressed before 
stable goals. Priorities represent standard practices 
of trauma care: if there are no differences in urgency, 
problems involving the airway are addressed before 
those involving breathing, which are addressed be- 
fore those involving circulation, etc. (the "ABCs of 
trauma care"). 
3 Interpreting and Reacting to 
Orders 
We have claimed that in patient-centered activities, 
the effectiveness of a "by-stander" decision-support 
system depends on focusing its clinical role on the 
immediate needs of physicians rather than interven- 
ing with information that, while correct and apro- 
pos, would not make a significant difference to pa- 
tient management. In order to achieve this, our 
approach is largely reactive rather than proactive. 
As noted in Section 1, instead of presenting Traum- 
AID's recommended management plan, TraumaTIQ 
waits for the physician's orders, which are treated 
as intentions to perform those actions. TraumaTIQ 
interprets and evaluates these intentions, and gen- 
erates a comment if it would be relevant to do so. 
TraumaTIQ also interprets the lack of an ordered 
action within an appropriate time frame as indicat- 
ing the physician's lack of intention to perform that 
action and evaluates it as such. This approach is a 
form of critiquing, early examples of which are (Lan- 
glotz and Shortliffe, 1983; Miller, 1986). 
TraumaTIQ's critiquing process is triggered 
whenever new information is entered by the scribe 
nurse and delivered to TraumAID. This information 
can be in the form of (1) bedside findings, (2) di- 
agnostic test results (indicating both that a diag- 
nostic action has been performed and what the re- 
sults of that action were), (3) therapeutic actions 
performed, or (4) diagnostic or therapeutic actions 
ordered by the physician. TraumaTIQ interprets the 
physician's orders in a goal-directed manner, using 
TraumAID's representation of goals, procedures and 
actions, so that the critique can address the likely 
reasons underlying any discrepancies and can sug- 
gest alternative means of addressing a particular 
goal. Figure 2 shows the architecture of Trauma- 
TIQ, comprising plan recognition, plan evaluation, 
and critique generation. 
3.1 Plan Recognition 
In order to judge when it is relevant to comment 
on a physicians orders or lack thereof, it is useful to 
consider not only what actions have been ordered, 
but also why. For example, this can allow comments 
to be withheld if an ordered action that differs from 
TraumAID's choice would nevertheless satisfy perti- 
nent management goals. TraumaTIQ's plan recog- 
nition component aims to infer the underlying goal 
structure motivating the physician's orders for ac- 
tions to be performed. Note that this is an example 
of keyhole recognition, rather than intended recogni- 
tion (Kautz, 1990). The physician does not intend to 
provide enough information to allow others to infer 
his plans. Like other members of the trauma team, 
TraumaTIQ must infer the plan incrementally based 
on what is known about the patient and what has 
been ordered so far. 
Trauma management poses particular problems 
for plan recognition. First, there is no fully pre- 
set sequence in which actions are ordered and done. 
Secondly, multiple diagnostic and therapeutic goals 
may simultaneously be active, and a single action 
can often be used to address several related or un- 
related goals. In this domain, it would thus be in- 
correct to simplify search by minimizing the number 
of top-level goals in the inferred plan, as Kautz and 
others have proposed (Kautz, 1990). Third, a plan 
recognition strategy cannot assume that the physi- 
cian's plan is correct: management of trauma pa- 
tients must often be done late at night and under 
time pressure, conditions that can lead to less than 
optimal decision-making even in trained profession- 
als. Since the number of possible incorrect plans is 
too large to encode a priori, we must make certain 
assumptions about physicians' plans: 
• The head of a trauma team will have expert or 
near-expert knowledge of trauma, and will usu- 
ally develop plans that are similar to Traum- 
AID's. Thus, if an action has been ordered 
that is also in TraumAID's plan, TraumaTIQ 
assumes that it is being done for the same rea- 
son(s). We call this giving the physician the 
"benefit of the doubt." 
• The physician is more likely to have appropriate 
goals but be addressing them in a sub-optimal 
way, than to be pursuing the wrong goals alto- 
gether. 
• While TraumAID follows a conservative strat- 
egy for pursuing diagnosis and treatment from 
observations, physicians may proceed more 
rapidly, pursuing a goal that may be involved 
in a current strategy but for which TraumAID 
does not yet have enough evidence to conclude 
its relevance. An understanding of the strate- 
gic relationships between goals should help to 
recognize examples of this difference. 
Reflecting these assumptions, the plan recognition 
algorithm works as follows: 
1. When an action, a, is ordered by the physician, 
check whether ~ is currently a part of Traum- 
AID's recommended plan as a means of satisfy- 
ing all or part of goal 7, or all or part of each 
member of a set of goals F. 
2. If so, add 3' or F to the representation of the 
physician's plan. 
3. If a is not currently in TraumAID's plan, de- 
termine whether there is a relevant goal that a 
might address: 
(a) If any of the goals that might lead to doing 
are present in TraumAID's current set of 
active goals, assume that c~ is being done 
to address that goal or goals. 
(b) In the case that there is no relevant goal 
to explain why the physician is ordering a, 
check whether any of the possible goals mo- 
tivating a are part of a currently active di- 
agnostic strategy (cf. Section 2). 
(c) If no relevant goal or strategy is found, 
leave the goal unspecified and add the in- 
tention to do a to the representation of 
the physician's plan with no goal attached. 
There is one exception to this rule: 
(d) If the system only knows of one possible 
goal that would lead to performing c~, Trau- 
maTIQ assumes that a is being done to 
address that goal, even though it does not 
consider the goal to be relevant. 
3.2 Plan Evaluation 
After attempting to infer the goal(s) underlying the 
physician's actions, TraumaTIQ attempts to identify 
whether commentary is warranted on any aspects 
of his plan. Plan evaluation begins by comparing 
the plan attributed to the physician with the plan 
produced by TraumAID, looking for four types of 
discrepancies: 
128 
TraumAtD 2.0 \] 
(~. patient "~ (f actions .~_..~ ~lan~ 
I Generation \[ ~ .... ~/ 
Figure 2: The TraumaTIQ module 
• Omission: A goal that TraumAID considers rel- 
evant is not being addressed by the physician 
in a timely manner. This can be further an- 
alyzed as to whether (1) the goal is not being 
addressed at all, or (2) the goal is only being 
partially addressed - some but not all the ac- 
tions in the procedure addressing the goal have 
been ordered. 
• Commission: An action is present in the physi- 
cian's plan that does not address a relevant goal. 
If a unique goal can be inferred to explain this 
discrepancy, that goal can be further catego- 
rized as to whether (1) it is unwarranted, (2) it 
is not fully proven, or (3) it has already been 
addressed. 
• Procedure choice: A relevant goal is being ad- 
dressed, but not using the procedure preferred 
by TraumAID. 
• Scheduling: Actions are not being done in the 
order recommended by Traumaid, e.g., satisfy- 
ing urgent goals before non-urgent ones. 
TraumaTIQ upholds the Maxim of Relevance by 
attempting to avoid comments that may later prove 
irrelevant. Errors of omission are not critiqued until 
a sufficient period of time has elapsed during which 
the physician might order the action. The amount 
of time allowed depends on the urgency of the goal 
that the action is intended to address - the rule of 
thumb TraumaTIQ uses is that a comment should 
be produced after approximately 10% of the time 
period has passed that is available to address the 
goal without significant consequences. Furthermore, 
comments are only made with respect to goals that 
cannot be made irrelevant by actions scheduled to 
be done before those goals are addressed. 
Discrepancies of all types are then evaluated in 
terms of their potential clinical significance. Cur- 
rently this is only a rough estimate based on ap- 
proximate cost and whether or not a procedure is 
invasive. For example, an unnecessary chest x-ray is 
considered insignificant while an unnecessary (and 
invasive) laparotomy is considered clinically signif- 
icant and worth drawing to the physician's atten- 
tion. We are currently in the process of developing 
more objective criteria for classifying errors accord- 
ing to their potential impact on the outcome of the 
case. Each discrepancy will be classified as either: 
(1) tolerable, probably harmless, (2) non-critical, but 
potentially harmful, or (3) critical, potentially fatal. 
Anything in the second or third category will be con- 
sidered significant enough to be reported in the cri- 
tique, while tolerable errors will not be mentioned. 
The output of plan evaluation is a set of com- 
municative goals containing (1) a propositional con- 
tent (PC) indicating the type of discrepancy and the 
particular TraumAID concepts involved, and (2) an 
illocutionary force (IF), such as URGE, INFOrtM or 
aEMIND, indicating how the information should be 
realized linguistically. For example, the goal: 
(SUGGEST 
(PROCEDURE-CHOICE GET-X-RAY-LAT- 
ABD 
GET-CT-SCAN-ABD 
129 
P~O-CoMPOUND-FRACTURE-LUMBAR- 
VERTESRA)) 
would be produced in a situation in which the physi- 
cian has ordered a CT-scan of the abdomen, which 
TraumaTIQ has inferred is intended to address the 
goal of diagnosing a compound fracture of the lum- 
bar vertebra, which TraumAID has instead chosen 
to address with a considerably less costly, less time- 
consuming lateral abdominal X-ray. The illocution- 
ary force of SUGGEST indicates that this is a non- 
critical error. Had it been evaluated as a critical 
error, the illocutionary force would have been URGE. 
3.3 Critique Generation 
Critique generation serves to organize TraumaTIQ's 
communicative goals according to the management 
goals they address and to translate them into Nat- 
ural Language utterances. This is currently the 
least sophisticated part of the process: each IF-PC 
pair indexes a sentential template with syntactically 
marked slots to be filled in with the appropriate 
phrasal translation of a TraumAID concept. For ex- 
ample, the template indexed by (SUGGEST, PROCE- 
DURE CHOICE) is: 
"TraumAID suggests (ARG1 GERUN- 
DIVE) rather than (ARG2 GERUN- 
DIVE). The former is preferred for (ARG3 
GERUNDIVE)." 
The first and second slots are filled in with gerun- 
dive phrases corresponding to GET-X-RAY-LAW- 
ABD and GET-CT-SCAN-ABD, while the third slot 
is filled in with an untensed verb phrase corre- 
sponding to R,O-COMPOUND-FRACTURE-LUMBAR- 
VERTEBRA, resulting in the sentence: 
"TraumAID suggests getting a lateral X- 
Ray of the abdomen rather than getting a 
CT-sean of the abdomen. The former is 
preferred for checking for fracture of the 
lumbar vertebrae." 
Critiques can be delivered directly to the trauma 
bay through synthesized speech or to the scribe 
nurse for subsequent delivery to the trauma team. 
4 Action Overloading 
As mentioned earlier, it is possible for a single action 
to participate in addressing more than one manage- 
ment goal. This strategy of doubling-up the use of 
plan elements is called intention overloading by Pol- 
lack (Pollack, 1991), who argues that it can simplify 
the process of computing optimal plans. It can, how- 
ever, complicate the process of recognizing and eval- 
uating them. Our strategy in TraumaTIQ is to as- 
sume that overloading occurs in the physician's plan 
whenever it would be beneficial. In other words, if 
an action ordered by the physician can be used to ad- 
dress two or more relevant goals, it is assumed that 
that is what the physician intends. This assump- 
tion makes sense because as long as both goals are 
being addressed, there is no need to produce a cri- 
tique, even if the physician did not have both goals 
in mind. 
If, on the other hand, there is a possibility for over- 
loading that the physician does not exploit, it may 
be clinically relevant to mention it. For example, 
consider a patient with a gunshot wound to the ab- 
domen, loss of sensation in both legs, and hematuria 
(Figure 3a). From the first two findings, TraumAID 
derives a goal of diagnosing a compound fracture of 
the lumbar vertebra, and from the type of wound 
and the finding of hematuria it derives a goal of di- 
agnosing renal injury. Both goals can be addressed 
by a single abdominal CT-scan. 
Figure 3b shows the plan that is inferred from the 
physician's order of a lateral abdominal x-ray, which 
is appropriate for the goal of diagnosing a fractured 
vertebra, but not for diagnosing renal injury. In this 
situation, two points are clinically relevant: (1) the 
goal of diagnosing renal injury should be addressed, 
and (2) both goals can be addressed with the single 
action of an abdominal CT-scan, obviating the need 
to do an X-ray to address one'of them. In this case a 
comment such as the following would be produced: 
"TraumAID suggests getting an abdominal 
CT-scan rather than a lateral abdominal 
X-ray. The former can be used both to di- 
agnose renal injury and to diagnose a com- 
pound fracture of the lumbar vertebrae." 
5 Conclusion 
We have presented an approach to information de- 
livery during patient-centered activities, in which a 
health care provider's (HCP) attention is on his or 
her patient and not on a computer screen. As well 
as potential applications to other areas of medical 
care, the idea of restricting the output of a decision- 
support system to clinically relevant information can 
be generalized to other task-centered activities, such 
as repair or construction, in which the person for 
whom the system output is intended is focused on 
performing a task rather than on the system. 
We recognize the need to validate the approach de- 
scribed here, and have proposed (with Sandra Car- 
berry, of the University of Delaware) such a study to 
the National Library of Medicine. In this study, we 
will compare two versions of our approach against 
graphic display of the stable portion of TraumAID's 
management plan on a monitor positioned in the 
trauma bay. In one version of our approach, cri- 
tiques of physician orders (or lack thereof) will be 
conveyed textually to the scribe nurse, who will con- 
vey it to the other members of the trauma team at 
his or her discretion. In the other version, critiques 
will be conveyed directly to the trauma bay through 
synthesized speech. 
130 
Hematuria Gunshot Wound Loss of Sensation 
in Abdomen in both legs 
I Rule Out I Rule Out Compound Fracture 
Renal Injury of Lumbar Vertebra 
Get CT-scan Get X-ray of 
Of Abdomen Lateral Abdomen 
CT-scan of 
Abdomen 
a) TraumAID's plan with overloading 
Rule out Compound Fracture 
of Lumbar Vertebra 
Get X-Ray of Lateral Abdomen 
X-Ray Lateral Abdomen 
b) Inferred plan after ordering X-ray 
Figure 3: Inferring a plan with overloading 
Even though this validation has not yet been done, 
we believe this paper has independent value in (1) 
pointing out additional constraints on language gen- 
eration raised by the desire to convey information to 
listeners attending to something other than an com- 
puter terminal, and (2) pointing out some features 
of plan inference and evaluation raised by multiple 
goal planning in a complex domain. 

References 

Clarke, J.R., Rymon, R., Webber, B., Hayward, C., 
Santora, T., Wagner, D. and Ruffin, A. The Im- 
portance of Planning in the Provision of Medical 
Care. Medical Decision Making 13(4), October- 
December 1993, p. 383 (abstract). 

Clarke, J.R., Webber, B., Niv, M., Rymon, R., Gert- 
ner, A. and Kaye, J. The Care of Injured Pa- 
tients: An architecture of Medical Knowledge. 
Der Chirurg (Special issue on surgical decision- 
making), to appear April 1994. 

Gertner, A. Ongoing Critiquing During Trauma 
Management. Proc. AAAI Spring Symposium on 
Artificial Intelligence in Medicine: Interpreting 
Clinical Data, Stanford CA, March 1994. 

Gertner, A. Responding to Users' Informational 
Needs in Time-Critical Situations Proc. 4th International Conference on User Modeling, Hyannis, 
MA, August 1994. 

Grice, H.P. Logic and Conversation. In H.P. Grice, 
Studies in the Way of Words. Cambridge MA: 
Harvard University Press, 1975. Reprinted in S. 
David (ed.) Pragmatics: A reader. Oxford: Ox- 
ford University Press, 1991, pp. 305-315, 

Kautz, H. A circumscriptive theory of plan recog- 
nition. In Jerry Morgan Philip R. Cohen and 
Martha E. Pollack, editors, Intentions in Com- 
munication. Bradford Books, 1990. 

Langlotz, C. P. and Shortliffe, E. H. Adapting a con- 
sultation system to critique user plans. Interna- 
tional Journal of Man-Machine Studies, 19:479- 
496, 1983. 

Miller, P. L. Expert Critiquing Systems: Practice- 
Based Medical Consultation by Computer. New 
York: Springer-Verlag, 1986. 

Pollack, M. Overloading Intentions for Efficient 
Practical Reasoning. N6us XXV(4), 1991, pp. 513- 
536. 

Rymon, R. Diagnostic Reasoning and Planning in 
Exploratory-Corrective Domains. PhD Disserta- 
tion. Dept. Computer & Information Science, U. 
Pennsylvania, November 1993. (Appears as Tech- 
nical Report MS-CIS-93-84). 

Rymon, R., Webber, B. L. and Clarke, J. 
R., Progressive Horizon Planning - Planning 
Exploratory-Corrective Behavior. IEEE Transac- 
tions on Systems, Man, and Cybernetics 23(6). 
Special issue on Planning, Scheduling and Con- 
trol, November 1993. 

Sperber, D. and Wilson, D. Relevance: communi- 
cation and cognition. Harvard University Press, 
Cambridge, Mass., 1986. 

Webber, B., Rymon, R. and Clarke, J.R. Flexible 
Support for Trauma Management through GoM- 
directed Reasoning and Planning. Artificial Intel- 
ligence in Medicine 4(2), April 1992, pp. 145-163. 
