Using ethnography in the design of an explanation system

Expert SystemsWithApplications,Vol. 8, No. 4, pp. 403-417, 1995 Copyright© 1995ElsevierScienceLtd Printed in the USA.All rights reserved 0957-4174/95 $9.50 + .00

Pergamon

0957-4174(94)E0032-P

Using Ethnography in the Design of an Explanation System DIANA E. FORSYTHE Intelligent Systems Laboratory,Department of Computer Science, University of Pittsburgh, Pittsburgh, PA

Abstract--In order for knowledge-based explanation systems to be acceptable, they must be useful and understandable to users. This implies that first, they should satisfy users' information needs and take account of their perspective; and second, they should be able to engage in dialogue with users. Much more progress has been made toward meeting the second condition than the first. Systems are still being produced that can engage in dialogue with users, but whose design reflects no systematic investigation of what people actually want (or need) to know about a given domain. Such lack of attention to the information needs of potential users is bound to limit the utility of any system. Because this issue relates to the concerns of social science as well as artificial intelligence, social scientists can help designers address it. One way of obtaining reliable data on the needs and characteristics of future users is ethnography, an anthropological method for gathering data in complex real-world settings. This article discusses some of the ways in which ethnography can contribute to the design process, drawing examples from an ongoing project to build an explanation system in migraine. Four aspects of our experience in using ethnography in the design process are discussed." rethinking basic design assumptions, investigating information needs, addressing the problem of perspective, and developing explanatory material. Based on this experience, the article concludes with the suggestion that the concept of explanation needs to be broadened still further to include more types of knowledge in the dialogue.

1. I N T R O D U C T I O N

broader and more complex (Swartout & Moore, 1993). In more recent work, explanation is viewed as a dialogue in which information is exchanged between the user and the system (Moore & Paris, 1992; Moore & Swartout, 1989, 1990). Such information exchange enables users to provide some information about their needs and preferences, as well as indicating (e.g., by selecting a particular word or passage on the screen) the source of the problem at hand. Much less progress has been made toward meeting the first condition. Explanation systems are still being produced that are capable of engaging in dialogue with users, but whose design apparently reflects no systematic investigation of either the context in which the system is intended to be used or of what people in that context actually want (or need) to know about the domain. Although system-builders may assume that they can accurately predict the information needs of future users and formulate material that will meet these needs, such untested assumptions are not a reliable basis for design (Forsythe & Buchanan, 1991; Wynn, 1991). This lack of attention to the information needs of potential users is bound to limit the utility of any system. For example, when designers rely on their own intuition concerning users' needs, they tend to design for themselves. In consequence, other users of the sys-

IN ORDERFOR know!edge-based explanation systems to be acceptable, they must be useful and understandable to users. We assume this means that first they should offer information that users want, presenting this materim in ways that make sense in terms of the users' own experience; and second, they should be able to engage in dialogue with users in order (among other things) to assess the extent to which their information needs have actually been met. Considerable progress has been made toward meeting the second condition. In the early days of expert systems, explanation (when present at all) entailed a one-way flow of information from system to user. It could not be adapted to the circumstances or information needs of particular users. For example, in MYCIN, an early expert system for medicine, " w h y " was always interpreted as a query about which rules had fired to produce a given output (Buchanan & Shorfliffe, 1984). Since this early work, the notion of explanation in artificial intelligence (AI) has evolved to become

Requests for reprints should be sent to Diana E. Forsythe, School of Engineering, Building 370, Standford University, Stanford, CA 94305. E-maih [email protected] 403

404

tem may encounter the problem of perspective, in which confusion arises because they do not share and may not even be aware of assumptions taken for granted by designers and thus embedded in their systems (Forsythe, 1993a; Forsythe, 1993b; Nyce & Bader, in press; Suchman, 1987; Suchman, 1990). Since these issues relate to the concerns of social science as well as AI, social scientists can help designers to address them. 1 One way of obtaining reliable data on the needs and characteristics of future users is ethnography, a research method developed by anthropologists to gather descriptive qualitative information in complex real-world settings (Werner & Schoepfle, 1987). The central concern of anthropology is investigating and documenting the perspectives and practices of different human groups, a task to which trained anthropologists bring specialized theory and methods. These skills have proven to be useful in the design and evaluation of knowledge-based systems (Fafchamps, Young, & Tang, 1991; Forsythe & Buchanan, 1991; Lundsgaarde, 1987; Nyce & Timpka, 1993). This article demonstrates some of the ways in which ethnography can help designers to understand users' information needs and to take account of their perspective(s). It draws examples from an ongoing project to build a patient education system with a sophisticated explanation capability. In 1991, we received funding from the National Library of Medicine (a division of the National Institutes of Health) to support an interdisciplinary project to design and build a patient education system using ethnographic data (in addition to other knowledge). The domain chosen was migraine. Now under construction, this system is designed to communicate in several ways with both doctors and patients. After taking a detailed initial patient history, it will formulate a written summary for the neurologist; following each patient encounter, it will provide both onscreen and written explanatory material that is tailored to individual patients. This material contains information about each patient's condition and current medications, offering further explanation on pertinent topics. In addition, it includes some general (i.e., nontailored) information about the experience and treatment of migraine that we believe to be relevant to all migraine sufferers--but which is seldom addressed explicitly by physicians. For further detail on this system, as well as the project, see Buchanan, Moore, Forsythe, Carenini, Banks, and Ohlsson (in press), Carenini and

Over the past decade or so, ideas from anthropology and sociology have been entering AI and cognitive science. A m o n g these is the importance of understanding ideas and practice in context. In AI, this notion has become identified with use of the term ' 'situated," a term borrowed from the American sociological tradition (see Mills, 1940) and first applied within AI by Suehman (1987). For discussion of this notion from various disciplinary standpoints, see the papers in Norman (1993).

D. E. Forsythe

Moore (1993a), Carenini and Moore (1993b), and Moore and Ohlsson (1992). Section 2 provides brief a overview of migraine and our use of anthropology in this project. Sections 3, 4, and 5 describe three interrelated aspects of our design process in which ethnography has played a central role: rethinking basic design assumptions, investigating the information needs of potential users, and addressing the problem of perspective. Section 6 illustrates how insights from ethnographic fieldwork have been translated into explanatory material to be generated by the system. The article concludes with a discussion that comments upon deriving design considerations from anthropological research, changing assumptions in midstream, the importance of information exchange, and the necessity of broadening (still further) our concept of explanation.

2. BACKGROUND 2.1. Migraine Migraine affects approximately 20% of the population (Lane, McLellan, & Baggoley, 1989), afflicting about three times as many women as men (Saper, Silberstein, Gordon, & Hamel, 1993, p. 93). Although rarely fatal, it causes considerable discomfort, reducing the quality of life of migraine sufferers and (in the United States) accounting for an estimated 74.2 million days per year of restricted activity (Stang & Osterhaus, 1993, p. 29). From the standpoint of physicians, migraine is sometimes difficult to diagnose and can be time-consuming and awkward to treat. In the diagnosis and treatment of this condition, effective information exchange between physician and patient is critical. Diagnosis depends primarily on a thorough history of symptoms: there are few physical signs or laboratory tests that differentiate migraine from cluster headache, tension headache, post-traumatic headache, or the somatic complaints of depression. A significant hindrance to the effective treatment of migraine is the fact that a substantial proportion of patients do not get sufficient relief from the regimen initially prescribed. Finding the proper medicine and dosage may require a long series of visits to the clinic; patients who become impatient or discouraged may be doomed to bounce from doctor to doctor without ever finding a successful treatment. In migraine, then, patient compliance is a major issue: patients must be motivated to return for further visits even when their treatment to date has been unsuccessful. These characteristics helped to make migraine a suitable domain for the research described here. Because migraine is diagnosed on the basis of patient history, doctor and patient must of necessity articulate much of the information exchanged, thus making it available to an observer as well. By the same token,

Using Ethnography in Design

since information that humans articulate can presumably also be transmitted (requested and provided) in a text format, we expected migraine to be a suitable domain for information transmission by means of a knowledge-based system.

2.2. The Role of Anthropology in the Project In this project, ethnography and an anthropological perspective contribute to system design in several ways. (A) First, to investigate the need for and the nature of explanation in migraine, the two anthropologists on the research team (D. Forsythe & M. Brostoff) have carried out extensive fieldwork. Observation of doctor-patient communication in a range of clinical settings has provided data about the information needs expressed by migraine patients, the explanations offered them by physicians, and the extent to which particular explanations satisfy particular information needs. To date we have carried out a total of 70 hours of field observations of doctor-patient visits in seven neurology-related sites in the Pittsburgh area. We have observed 78 neurology patients, of whom 29 suffered from headaches. The providers observed include seven attending neurologists, six residents, five medical students, and four other health care providers. Of the doctor-patient encounters, 36 were tape recorded and have been transcribed. In order to supplement data gained from observations and to investigate information needs that patients might not be expressing to their physicians, we have ufidertaken a series of semidirected interviews with migraine sufferers outside the clinical setting. These interviews average a little over 1 hour in length, and have provided detailed data on matters that have seldom arisen during our observations of patient visits. To date, we have interviewed four men and eight women, producing 12 hours of tape-recorded material, which has also been transcribed. In addition to these relatively formal taped interviews, we have carried out several dozen informal interviews with migraine sufferers encountered in the course of the ongoing fieldwork. These informal interviews are recorded in handwritten field notes but not tape recorded. (B) Second, we take an interpretive anthropological approach (Geertz, 1973, 1983; Traweek, 1993) to analyzing the transcribed texts of observations and interviews. Content analysis of these transcripts identifies information needs expressed by patients and physicians, explanations offered by physicians, and data about how these explanations are received. It also provides insight into the types of knowledge physicians and patients have about migraine and its treatment, as well as what each takes for granted in discussing issues and procedures in this domain. Following analysis of individual texts, we compare them, seeking patterns and common features within and between them. This

405

process provides insights for both the content and wording of explanatory material to be offered by the system. (C) Third, as the system is constructed, the project anthropologists draw upon their field experience and their understanding of the domain to help insure the appropriateness of the interfaces and the explanatory material developed by other team members. The anthropologists also develop explanatory material of their own to be integrated into the text generated by the system (examples are given below). (D) And finally, the anthropologists have played a reflexive role in the project, reflecting back to their fellow researchers some of the design teams' own tacit assumptions about the nature of their task. Several of these assumptions appeared inappropriate in the light of our field data; in consequence, they have been modified (see below). As the research progresses, fieldwork, analysis, and system design proceed on an iterative basis. Having conducted a period of fieldwork, we review the data in search of patterns, construct preliminary concepts and inferences, reflect upon their system-building implications, and then return to the field to further investigate and test these ideas. Concerns of system design sometimes suggest specific issues for the fieldworkers to investigate (e.g., the way causal relations are explained); similarly, the fieldwork~reveals issues with implications for the system (e.g., patient fears). The interrelated processes of fieldwork and system design will continue throughout the project. However, we have already learned some important things from the effort to incorporate ethnography into the systembuilding process. This article presents some of what we have learned.

3. RETHINKING BASIC DESIGN ASSUMPTIONS Medical information systems necessarily incorporate assumptions about the world, some of which are explicit and some of which remain tacit (Kaplan, 1987). Traditionally, the information built into a system stems from one of two sources: what the designer knows and/or takes for granted about the world (Forsythe & Buchanan, 1991), and the conscious models of the domain expert(s) involved in the design team (Forsythe, 1993a). Both types of belief can influence decisions about system design, although they may not have been subjected to rigorous scrutiny. Incorporating ethnography into the design process can help to reduce the extent to which design is based upon unexamined assumptions. Given a system intended for a particular population of users, we can investigate systematically the information needs of such users, making informed inferences about the characteristics that a system would need to possess in order

406

to meet those needs. Real-world data can thus assist designers to evaluate the appropriateness of their own prior beliefs about what potential users want and need. In the first year of the project, our fieldwork in neurology led us to reassess some of the beliefs upon which the migraine project had been based, Initially, we had made a number of assumptions concerning the information needs of migraine patients, the meaning of such concepts as "tailoring" and "explanation," and the factors that influence successful information transmission. Some of these beliefs were explicit, while others only became clear as the fieldwork made us aware of other ways of thinking about these issues. To illustrate this process, this section considers five assumptions and the way in which the ethnography has led us to rethink them. 1. At the start of the project, we believed (and doctors told us) that patients actively request explanatory material from their physicians. We also believed (and doctors told us) that physicians respond to such requests by providing extensive explanatory material. This was the rationale for gathering data by observing doctor-patient communication: we assumed that watching this process would provide useful data for building a patient education system. 2. Referring to patient education in diabetes, Juge (a physician) notes that "patient education is often limited to teaching physiopathology of diabetes, or even biochemistry" (Juge & Assal, 1992, p. 217). Initially, adopting our expert neurologists' point of view, we made a parallel assumption, taking for granted that the explanation that migraine patients want is textbook material concerning the physiology of migraine and the side effects of particular drugs used to treat it. 3. Because we assumed that migraine patients' information needs result from a lack of formal, general information 2 about migraine, we expected to build a conventional, intelligent tutoring system. 4. Believing (as we still do) that it is useful to tailor explanation to individual patients, we conceived of such tailoring as adapting the content of this textbook information to individual patients. Related to this, we assumed that the "patient types" to which we would adapt this information would be defined in terms of patients' social attributes (e.g., ethnic background, educational level, age, and possibly gender.)

2In an earlier article on physicians' informationneeds in internal medicine, we pointed out that medical informationvaries both in its degree of specificity and in its degree of formality. Crossing these two dimensions produces a matrix of types of medical information (Forsythe, Buchanan, Osheroff, & Miller, 1992). Most knowledgebased systems, including those in medical informatics and tutoring, are designed to transmit formal, general information--that is, information of the sort found in textbooks.

D: E. Forsythe

5. Following from Assumption 2, we believed that patients would be empowered and compliance would be increased if we simply provided them with more and better substantive information about the physiology and treatment of migraine and the side effects of migraine drugs. On the basis of these assumptions, we began empirical research to investigate just what explanatory material migraine patients request of their physicians and to identify the "patient types" in terms of which this material should be tailored. At the same time, the system-builders began to develop schemes for representing textbook material on the physiology of migraine and the side effects of particular migraine drugs. However, as fieldwork progressed, the anthropologists on the design team began to have doubts about these assumptions.

3.1. Assumptions

3.1.1. Assumption 1. The patients we observe rarely spontaneously demand explanations from their doctors. Patients don't ask nearly as much as we thought they would, and physicians don't necessarily give long or coherent explanations. This is consistent with the literature on doctor-patient communication (Frankel, 1989; Wallen, Waitzkin, & Stoeckle, as cited in West, 1984, p. 108), but contradicts what we had been told by some physicians. In order to elicit information needs and other concerns that patients might not be expressing to their physicians, we began a series of semidirected interviews with migraineurs outside the clinical setting. These interviews have revealed that patients have important information needs that they do not necessarily bring up with their doctors. The interviews provide extensive data on matters that have seldom arisen (or have arisen only obliquely) during the patient visits that we have observed (see below). 3.1.2. Assumption 2. As fieldwork got underway, we discovered that while migraine patients do sometimes have questions about the physiology of migraine and the side effects of different drug treatments, they also have numerous other concerns. We realized that we had simply been assuming that what patients want to hear about migraine is what doctors want to explain. In actuality, patients want to know about a broad range of topics involving migraine. These concerns can be categorized as follows: 1. Is this something major (e.g., a stroke or brain tumor)? 2. How does this make sense in terms of what I know about my body and my life (e.g., my eating habits or menstrual cycle)?

Using Ethnography in Design 3. What will I need to do about t h i s - - h o w will it affect my everyday life (e.g., my work schedule or my ability to drive)? 4. Why is this happening to me (allergy? inheritance?)? This list reflects a hierarchy of concerns, with (1) clearly perceived as the most important. At the bottom of this list come concerns such as the following: 5. What is migraine anyway, and how does m y migraine drug work? 3.1.3. Assumption 3. The fieldwork has shown us that much of what patients want to know is informal and/ or specific knowledge, which may vary according to context and person. Our system needs to be able to convey this type of information as well as more formal, general knowledge of the sort found in textbooks. In addition, the system is intended to be used by adult migraine patients rather than children. In relation to physicians, the adult migraine sufferers we have observed and interviewed are best described as having alternative knowledge rather than a lack of knowledge about migraine. Both these factors led us to realize that the system under construction needed to b e something other than a conventional tutoring system. Our educational task is less to provide information to those who completely tack it, than to support the exchange of information between physicians and migraine sufferers. Of particular importance is helping individual patients to translate formal medical knowledge about migraine and its treatment into information that they can apply in their own lives, a factor that seems to promote patient compliance (Hunt, Jordan; Irwin, & Browner, 1989b). 3.1.4. Assumption 4. While we continue to believe that tailoring explanation to individual patients is desirable, the fieldwork has changed our ideas about how this should be done. Our notion of " p a t i e n t t y p e s " has evolved f r o m one defined by relatively stable social attributes to one that also reflects individual patients' current concerns with respect to their condition. Different patients have different information needs, and the needs of particular patients vary over time. In appropriate contexts, explanation generated by the system will need to be adapted to patients' gender and age, 3 and possibly to their educational level and ethnic background as well. However, it will also need to be tailored to patients' emotional

3 In women, migraine is often affected by the hormonal changes associated with events in the reproductive cycle (menarche, menstruation, pregnancy, menopause). It may also be affected by hormonal medications such as birth control pills and estrogen replacement therapy. Thus, explanatory material for women needs to be tailored by age; that for m e n does not.

407 concerns, including where individuals are in their understanding of their condition. Note that the list of patient concerns (given in section 3.1.2.) begins with specific, personal issues and moves toward more general ones that nevertheless still are personal. The order of these concerns is important: patients are unlikely to take in the later information (4 and 5) until they have received answers to the prior concerns (1, 2, and 3). For example, a lecture on the physiology of migraine or the side effects of a given drug (5) will not register with a patient worried about a possible brain tumor (1). Patients' information needs seem to start with the self and with major life/death worries, and to work from there outward; formal, general concerns come at the bottom of the list. Thus, textbook descriptions of migraine are of much less immediate concern to many patients than "what is going to happen to m e . "

3.1.5. Assumption 5. We initially saw tailoring explanation as a way to increase patient compliance, assuming that compliance will increase if information flow increases. For this reason, our initial discussion of tailoring explanation focused entirely upon the content of the explanatory material presented by the system. But the fieldwork shows that the effectiveness of physicians' explanatorY messages depends in part on the style in which material is presented: this reflects both the tone with which doctors approach their patients, and the way in which doctors frame what they say to them. In other words, how doctors say what they say is at least as important as what they say to patients. This led us to pay closer attention to questions of wording and style in developing explanatory material to be generated by the system.

3.2. Considering Explanation in Practice In real-world contexts other factors, in addition to formal content, affect whether or not an explanation is successful. A fact-flied explanation that is delivered in a way that a given patient cannot " h e a r " is not a good explanation in reality. Some of the factors of relevance here have to do with nonverbal aspects of the physician's self-presentation, such as tone of voice, gaze, and touch. These do not correspond in any obvious way to attributes of a knowledge-based system. But other factors do suggest possible design considerations for system building. For example, the field research led us to focus (inter alia) upon the process we call enlistment. Although we assume that all of the physicians we observe offer explanations that are technically correct, some doctors consistently word their explanations in ways that seem to " g e t through" more successfully to patients. One relevant factor that we have identified in this connection is the language in terms of which physicians frame

408

their questions and explanations. Enlistment of the patient by the physician is promoted by the use of inclusive language, that is, language that explicitly recognizes patients' competence and thus treats them with respect. We believe that such language increases the physician's (and possibly also the prospective system's) chances of having the patient hear the diagnosis and treatment recommendations. Given the importance of the doctor-patient relationship for patient education (Berg, 1987), as well as the links between d o c t o r patient communication, patient satisfaction, and compliance (Ley, 1982), it seems likely that successful enlistment of the patient by the doctor through the use of such language helps to promote compliance. The following examples illustrate the use of inclusive language for patient enlistment, a strategy that can certainly be used in language generated by a system. Text in italics is verbatim material spoken by physicians; the text in Example 3 summarizes the physician's utterances as recorded in the fieldnotes, but is not verbatim. 3.2.1. E x a m p l e 1. I'll take a look in y o u r ears, j u s t to be sure. A s you know, y o u r ears are connected to y o u r throat, and sometimes a sore throat can cause an ear infection. 3.2.2. E x a m p l e 2. The chances are pretty good that this is a virus, the flu. A s you know, there are two kinds o f germs, virus and bacteria (further explanation follows). These explanations contain some information about anatomy and germ theory. However, in each case, that information is presented after the phrase " y o u k n o w , " a phrase used quite consistently by this particular doctor. This wording allows the doctor to give patients information that they m a y or may not know without appearing to lecture them. This wording is inclusive because it treats the patient with respect, as someone who shares knowledge with the doctor. The explanation is presented as a reminder rather than as instruction. 3.2.3. E x a m p l e 3. The patient is a nurse with occasional migraines who has come to the doctor because of recently developed tension headaches. The narcotics that have worked occasionally in the past for the patient' s migraines are not working now that she is taking a lot of them for the continuous tension headache. The physician urges the patient to get off the narcotics, saying approximately: As a nurse you know all about drugs and have easy access to them. But we medical people need to watch out and realize that we too are subject to some of the problems with these drugs. In this example, a different physician uses the fact that this patient is a nurse to frame the message as a reminder between colleagues. The device " w e medical

D. E. Forsythe

people" is used to make more palatable the presumably unwelcome message that the patient has become habituated to the narcotics she is taking for migraine. 3.3.

Commentary

Our use of ethnography called into question some of the basic assumptions on which this project was based. Initially, the fieldworkers set out to discover the relation between migraine patients' information needs and their social and educational characteristics. Conceiving of these characteristics as defning patient " t y p e s , " we assumed that these " t y p e s " would provide an appropriate framework in terms of which to tailor educational information about migraine and drug side effects. This approach reflected the belief that our goal was to construct a tutoring system to convey in appropriately tailored language some version of the same basic textbook material to all migraine patients. This framing of the problem assumed the patients' information needs to result from a lack of formal, general information about migraine, that is, precisely the sort of knowledge in which physicians are trained and which intelligent tutoring systems are designed to teach. Analysis of our first year of field data led project members to rethink this approach. First, although migraine patients exhibit a range of social and educational backgrounds, it remains unclear at this stage whether there is a relationship between these characteristics and patients' information needs with respect to migraine. Thus, our initial notion of patient " t y p e s " needed to be modified. Second, patients do indeed express information needs with respect to migraine, but these needs do not seem to frequently involve textbook-type information. In other words, what physicians and cognitive scientists define as " e x p e r t i s e " - - f o r m a l , general information about anatomy, physiology, and pharmacolo g y - i s not necessarily what patients want to know. Instead, the primary information need of the migraineurs we observe and interview seems to be to discover some meaning in their condition and to understand the implications of that condition for their own lives. As is illustrated in the hierarchy of concerns presented above (see 3.1.2. Assumption 2), the "toplevel" question for migraine patients is, " w h a t will happen to m e ? " Because migraine patients often fear that they have a brain tumor or have had a stroke (see below), this is understood as a question of life or death significance. Lower down on the hierarchy are questions about the triggers, treatments, and causes of the condition. Useful responses to these questions may make use of textbook information, but will need to focus upon helping patients to translate this information into implications appropriate to their own particular circumstances. We hypothesize that every migraine

Using Ethnography in Design

patient perceives something like this hierarchy of concerns. We do not yet know how long it takes patients to move through the hierarchy, or the extent to which individuals vary in this progress. However, each step of this hierarchy is associated with a particular set of information needs. In offering information to migraine patients, then, we need to provide explanation that meets the information needs that they currently experience. This implies that instead of tailoring information in terms of patients' social characteristics, we should tailor that information to their current concerns about migraine. In a sense this makes our task more difficult: although at least some of patients' social attributes remain fairly stable over time, their concerns about migraine are bound to change as they move through the hierarchy. On the other hand. tailoring explanation to what patients are currently experiencing makes it much more likely that the system will actually help to meet their information needs. 4. I N V E S T I G A T I N G I N F O R M A T I O N NEEDS Designing an explanation system requires decisions about what to explain. Since the primary purpose of the system under development is to provide information for the benefit of migraine sufferers, a central goal of the ethnographic fieldwork is to investigate what their information needs actually are. The previous section touched upon this in extremely broad terms, in order to emphasize the necessity of addressing frightening top-level questions before providing material about other issues. This section takes up the question in more detail. People with migraine have many information needs. Some are made explicit during visits to the doctor; others remain partly or entirely tacit in the course of doctor-patient communication. Often such tacit information needs remain unsatisfied. For example, early in the study, one anthropologist noted during a patient visit that in speaking to the doctor, the patient made repeated but fleeting references to death, which the doctor seemed to overlook. Here is one such reference from the transcript of this visit:

4.1. Example 4 And when you get one of those (very bad migraines), what do you do ? P a t i e n t : Make sure the sheets are all nice and smooth... Neurologist: (Laughs) P a t i e n t : . . . nice and dark and pull the blankets up and die. J u s t . . . Neurologist: (Interrupts) You don't take any medicine for them? Neurologist:

(Transcript of 22 Nov. 91, Lines 109-118)

409

Intrigued, we examined the transcripts of other observation sessions, and realized that something similar had also happened on other occasions involving other patients. Hypothesizing that this use of death imagery reflected a significant underlying concern, we added some open-ended questions on this theme to our interview schedule. The interviews revealed that migraineurs do indeed often fear that their condition may be fatal, a concern apparently prompted by the frightening nature of some migraine symptoms and by their perceived similarity to symptoms of brain tumor or stroke (c.f., Good, 1992, p. 62). This example illustrates the flexibility of ethnography as a research method, an important aspect of its utility for system design. Having detected hints of some underlying fear from our observational data, we used interviews to pursue the matter, obtaining data that confirmed this insight and expanded our understanding of patient experience. Even when patients do bring up matters explicitly with their doctors, they always ask far fewer questions than the physicians. This can be seen in the ratio of information needs recorded in our transcri/pts of doctor-patient discourse. In the transcri/pl excerpted above, for example, the doctor expressed 127 information needs while the patient expressed 7 (22 Nov. 91); in another, the doctor expressed 39 information needs while the patient expressed 4 (10 Jan. 1992). One of the most verbal patients we observed managed to express 24 information needs during a single visit; during the same time span, however, her physician expressed 69 (23 July 1992). Observing patient visits did produce useful data on the information needs of migraine patients, but because of this communicative imbalance the interviews with migraineurs were a much richer source.

The transcript analysis underway will produce descriptive statistics on all information-seeking messages expressed by either patients or physicians during the patient visits observed. To provide more immediate design input, we have compiled a list of 166 information needs that migraine sufferers have expressed during patient visits or interviews. These questions are organized in the following categories: (a) what is a migraine; (b) general causes of migraine; (c) how lifestyle can affect migraine; (d) migraine triggers; (e) physiology, experience, and medical implications of migraine; (f) drug treatment for migraine; and (g) other questions about treatment. (Sample questions are given below.) Since both fieldwork and analysis continue, the list is still growing. Because spontaneously-expressed information needs are not always verbalized as syntactic questions (Forsythe, et al., 1992), some of the questions on the list are verbatim, while others are reformulations of utterances that conveyed an explicit desire for information. All are grounded in documented expressions of information-seeking by migraine suffer-

410 ers. While this list is not exhaustive, it serves as a useful resource to help the project team understand what potential users of our system are likely to want to know. Based on their understanding of the importance of these queries to those who posed them, the fieldworkers attempt to order them in terms of priority. Some questions on the list correspond to queries that neurologists told us to expect from patients. These are the sorts of questions that we anticipated providing answers for at the beginning of the project. Many of these are questions that we ourselves have observed patients asking. Four examples follow: Q1. What is a migraine? Q2. Are everyone's migraines the same? Q3. Does everyone have the same migraine triggers? Q4. What is the best medicine to take f o r migraine ? However, there are other questions on our list that doctors did not tell us to expect. We have not observed patients asking these questions of their doctors directly, although occasionally we note indirect reference to such concerns. Many of them were gathered and/or made clearer to us in the course of private interviews with migraine sufferers. The existence of such questions demonstrates the utility of collecting direct data on patients' information needs: physicians' models of what their patients want to know are unavoidably partial. Not only do patients in some cases fail to bring up concerns, but some of their information needs derive from differences in tacit knowledge between doctors and patients. Thus, patients may be puzzled by things that doctors are not aware they do not know. This problem is illustrated in the next group of questions. Here, each speaker is missing a piece of information that is probably obvious to neurologists but not necessarily to patients: Q5. Each time I get a migraine, I'm afraid I'm going to die. Could that happen ? Q6. My headaches hurt so much that I am afraid I have a brain tumor. Do I? QT. When I get a migraine, strange things happen to my vision. CouM my eyes be damaged or could I go blind because I have migraine ? Q8. My doctor told me that I have "atypical migraine." Is that worse than regular migraine? Neurologists know that migraine is rarely fatal (Q5). However, patients do not necessarily know this. Some equate their symptoms with those of brain tumor (Q6), stroke, or retinal detachment (Q7). Finally, the speaker in Q8 has an information need deriving from an unexplained label used by a physician, presumably left undeciphered because the term is so familiar to practitioners. In another group of questions, derived from interviews, migraine sufferers express concerns about the actions or inaction of physicians:

D. E. Forsythe Q9. I think that I can get migraines from something I ate, but my doctor doesn't seem to take that very seriously. Is he right? QIO, I had migraine f o r 30 years before anyone told me about ergotamine. Why would doctors withhold that information from me? Q l l . I've been to lots o f doctors but no one has cured my migraines. I don't know if they didn't believe me or if they were incompetent. It is difficult to know what the doctors' point of view was in Q9 and Q10, but clearly these patients would have benefited from more extensive explanation about their condition and its treatment. Q11 suggests that the patient lacks two pieces of knowledge about medicine in general and about migraine treatment in specific that neurologists take for granted. First, many cases of migraine can be controlled but not cured; that is, the pain and/or frequency of attacks may be reduced but not completely eliminated. Second, it often takes trial and error to determine the most suitable medication and dosage for each patient. Thus, patients need to return to the physician repeatedly, even when successive treatments have been unsuccessful. Finally, with Q12, we illustrate a further type of information need that is virtually universal in migraine, but in our experience rarely addressed by physicians. This is the problem of trying to manage real-world jobs and lives with a chronic condition that causes unpredictable, sometimes debilitating pain. Coping is made more difficult by the fact our society sometimes treats migraine as stigmatizing (as the speaker here suggests). Q12. As a professional woman, I wonder whether people will see my migraine as some kind of illness that could be considered a weakness. This is a brief sample of the information needs collected in the course of our fieldwork. A systematic analysis of our observational and interview transcripts is in progress. For the purpose of system-development, collecting and analyzing such data provides a pointer to topics that the explanation system should address. Section 6 provides examples of explanatory material developed to meet some of these concerns. First, however, Section 5 raises the issue of perspective.

5. A D D R E S S I N G THE P R O B L E M OF PERSPECTIVE

Patients do not necessarily view their condition in the same way their doctors do (Hunt, Jordan, & Irwin, 1989a; Kleinman, Eisenberg, & Good, 1978). This difference, which may affect compliance (Hunt et al., 1989b), raises difficult issues for patient education. For example, should explanation be driven by what patients want to know or by what providers believe they need to know? These may not be the same. While

Using Ethnography in Design

providers are likely to assume the latter, patients may prefer the f o r m e r - - b u t thereby miss information that their physicians see as important. The problem of perspective is brought into sharp relief by the task of designing a computerized explanation system. Face-to-face, humans can exchange unplanned messages, spontaneously adapting their explanations to individual listeners. In contrast, although it may be somewhat adaptable, computer-generated explanation has to be designed beforehand and must generate text that is appropriate and meaningful to many different users. This places great importance on decisions about what to include in the knowledge base. To date, few computerized patient education systems have been adopted into routine use in health care settings, perhaps in part because the technology is new. However. we believe that a nontechnical contributing factor is widespread failure during system construction to consider explicitly the issue of perspective. Sometimes system-builders appear to simply take it for granted that their perspective will be shared by future users; when systems are not accepted, users are blamed (Forsythe, 1992). The potential cost of this approach is considerable. Well aware of this history, we are attempting to take a different approach, using ethnography to incorporate information on the perspectives of both migraine patients and neurologists into the design of the system. Neurologists and migraine sufferers can both be seen as experts on migraine. However, what they k n o w - - a n d seek to know about migraine is somewhat different. Migraine sufferers put together much experiential knowledge about their condition. They know in detail what happens to them personally during a migraine, how it feels, what helps, and what doesn't help. This experiential knowledge is not necessarily compiled in terms of formal categories of diseases, symptoms, and side effects. Thus, to take an example from our fieldwork, a patient may not pay attention to whether dizziness experienced in the course of an attack is a function of the migraine or of a medication used to treat it. Rather. he may think of dizziness as simply something that happens when he gets migraines. Following the knowledge typology developed in our previous work (Forsythe et al., 1992), migraine sufferers have a large store of informal, specific knowledge about their own history and experience of migraine that tends to be understood as a whole, as " m y experience of migraine." In addition, since people with this condition often share knowledge with each other, migraine sufferers may also possess informal knowledge that is more generally applicable. In contrast, neurologists command a store of formal, general knowledge about migraine and its treatments information ordered in terms of categories of diseases, symptoms, drugs, and side effects accepted and taught throughout Western medicine. In order for

411

them to apply this knowledge to diagnose and treat migraines, it is necessary for them to relate these categories to what they hear from their patients. For purposes of medical decision-making, therefore, a neurologist may need to know whether a patient's dizziness results from his migraines or from a medication, or whether another patient's confusion stems from her headaches or from neurological damage caused by a car accident. (These examples are taken from our field data.) Neurologists also know about the practice of medicine, about the experimentation necessary to find the optimal treatment and dosage for individual patients, and about what sorts of treatment and pain relief are actually possible. This constitutes both formal and informal knowledge. Finally, they have considerable formal, specific knowledge about past and present patients in the form of medical records, test results, referral letters, etc. 4 In short, patients and their doctors have somewhat different kinds of knowledge about migraine and tend to stand in a different relation to that knowledge. Thus, even when physicians and patients address the " s a m e " topic, what they are talking about may not be the same. While the knowledge of both is relevant to the diagnosis and management of migraine, observation shows that the difference in their perspectives can impede the transmission of their knowledge to each other. Sometimes misunderstandings occur with respect to information that is expressed in the course of a patient visit. For example, a patient may not understand a term used by a physician or may not be able to provide information that is requested. In addition, both doctors and patients bring knowledge to their encounters that is not expressed. This too can contribute to misunderstandings between them. Although conversational partners often assume that others share their point of view (Tannen, 1986), it is evident to an observer that doctors and patients sometimes take quite different things for granted. This can cause them to miscommunicate. Some examples from our fieldwork illustrate this point.

5.1. Example 5 Doctor: What brought you here today? Patient: I came in a car. (Exchange reported to us by a physician) The physician here asks one question; the patient answers another. The question and answer fail to match because they reflect different assumptions: the patient interprets the query literally when it is not so intended.

4 For more detailed discussion of the nature of knowledge in neurology; see Nyce & Graves (1990).

412

D. E. Forsythe

What makes her response incorrect from the doctor's standpoint is the taken-for-granted cultural information that in this setting, he must be asking about her medical condition. (In another setting the patient's answer might be a reasonable one.) The patient in this example is mildly retarded and apparently lacks the tacit knowledge that makes the intended meaning of this query obvious to both the doctor and the reader. While this is a particularly clear example, however, this type of misunderstanding is by no means unusual. We have recorded many instances in which doctors and patients misinterpreted each other because they took different things for granted. Example 4 provides an illustration from the migraine study.

5.2. Example 6 Neurologist: How did you take the Fiorinal ? Patient: Caplets. Neurologist: But I mean every day you took them? Patient: Every day. Neurologist: Fiorinal every day? Are you sure it wasn't propranolol ?

(Transcript of 7 Feb. 92, Lines 94-102)

Here the initial question and answer fail to match because they reflect different assumptions. The patient interprets " h o w " to refer to the form in which she takes the drug, rather than the frequency. Her response reveals that she doesn't know what distinctions are likely to be of significance to a neurologist making treatment decisions for migraine. In this example, the misunderstanding is straightforward; the physician instantly perceives and corrects it. Sometimes, however, the nature of the misunderstanding is more complex and may not be repaired. A fertile source of miscommunication is the fact that busy physicians make tacit assumptions that are not always explained to patients. In some cases these assumptions seem idiosyncratic. For example, we observed a physician who apparently avoids prescribing for female patients a migraine drug that can cause weight gain. Patients may be confused as he mentions this drug as a treatment possibility and then dismisses it without explanation. Other tacit assumptions made by physicians concern accepted practice in neurology, which may be unknown to patients. For example, we observed a patient trying to understand why a physician was pressing her to try an abortive drug that (as he explained graphically) can cause acute nausea instead of prescribing a prophylactic medication that is not nausea-inducing. In discussion with the physician afterward, the fieldworkers learned that he was fol-

lowing an accepted treatment agenda: in patients with relatively infrequent migraines, abortive drugs are ruled out before prophylactic medications are tried. However. this agenda was not explained to the patient. Because she was unaware of something he took for granted, she had difficulty accepting his recommendation for treatment. An explanation of this piece of tacit medical knowledge could have helped to improve information exchange between doctor and patient. In a second illustration of this point, we observed a different patient who had been prescribed two drugs for her migraine: one to abort the headaches and an antiemetic to counteract the nausea-inducing properties of the first drag. However. although the prescribing physician obviously intended the drags to be taken together, this crucial piece of underlying knowledge was not made clear to the patient. She took the first drug on its own, became very sick, and gave up the drag without ever having taken it as the physician intended she should. While it may be obvious to a neurologist that migraine drugs prescribed together are meant to be taken at the same time, such background information is not necessarily obvious to patients. Like all tacit knowledge, it needs to be stated explicitly to be understood by those whose assumptions may be different. A different aspect of the disparity in perspective between migraine sufferers and neurologists is illustrated by the experience of patients who attempt to raise information needs with their doctors that lie outside the realm of what physicians tend to think of as medical knowledge. These are problems of the sort referred to in Q 5 - Q 8 and Q12 above; they include queries about how to deal with chronic pain, expressions of desperation by patients who feel that they cannot cope any more, and hints concerning fears of stroke, brain tumor, or death. Such questions are far removed from neurologists' preferred explanatory material on the pathophysiology of migraine and the side effects of drugs used to treat it. However, these questions communicate considerable emotional pain and can be very important to patients. In the encounters we have observed, physicians have for the most part not addressed such information needs. Often the observer has the impression that doctor and patient are talking past each other. For patients, having a physician fail to respond to concerns of this sort can be a painful and alienating experience; for their part, physicians may wonder why patients bring up issues that they don't really know how to address. To the observer, such incidents are a powerful reminder that physicians and patients have different perspectives on migraine, including somewhat different notions about what doctors should be asked to explain.

6. DEVELOPING EXPLANATORY MATERIAL As we have seen, ethnographic fieldwork has helped to reveal the information needs of migraine patients.

Using Ethnography in Design

It also has helped to make explicit some of what both physicians and patients take for granted about migraine. In analyzing the field data, we focus upon exploring its implications for the development of explan2 atory material for the system. The data have provided numerous insights about both what the system should say, and where and how it should say it. (It has also provided some insights about things the system should not say.) This section provides two examples to illustrate the process of translation from ethnography to explanatory text. Note that in each case, the text itself is fairly simple; however, it required insight to realize that explanation on these topics was required.

6.1. Example 7 As we have seen (see above), many migraine patients secretly fear that their attacks may be fatal. However, because they rarely, if ever, bring up this concern explicitly with their doctors, they never really have it allayed. (Several interviewees commented that while they obviously had survived each migraine attack to date, the pain of a severe attack was such as to lead them to fear that this one might kill them. Thus, past experience was never sufficient to dispel this particularly horrible private fear.) Having understood the nature of this concern, we decided that the system ought to address it directly--although we have never seen a physician do so. Among patients' concerns, the issue of whether their condition may be fatal is clearly at the top. It must be settled in order to enable patients to assimilate important but less emotionally compelling details about medications, lifestyle changes, and so forth. Since our interviews suggest that the fears described above are nearly universal among migraineurs, it makes sense to address them in a w a y that will reach every patient who uses the system. We decided that a message on this theme should be included in the interactive information sheet to be shown to all users of the system following their initial visit with the neurologist. This is described in some detail in (Buchanan et al., in press). Briefly, the interactive information sheet will first be presented on-screen. It will be tailored to individual patients and contain mouse-sensitive areas allowing users to query topics and ask follow-up questions. After each session with the explanation system, users will receive a printout of the interactive information sheet they just worked with; this will include a copy of the questions they just asked and the answers received. The following piece of text will be included in the interactive information sheet: While migraine can produce very intense and painful headaches, they are not life threatening. Sometimes patients worry that their migraine symptoms mean that they have a

413

brain tumor or have suffered a strokel Do not worry. You were diagnosed as suffering from migraine only. [*Your physical examination was normal, which indicates that more serious causes of headache such as bleeding or tumors are very unlikely.*] The first four sentences in this message are entirely canned; they will be seen by every patient who uses the system for the first time. The last sentence (in square brackets) will be included as appropriate when indicated by the physician. Clearly, this particular text offers only minimal potential for tailoring to the individual user. Note, however, that the use of largely canned text to address this initial general concern helps to prepare the ground for text later on in the interactive information sheet that is tailored on the basis of such factors as the user's gender, age, migraine triggers, treatment history, and current prescription (Buchanan et al., in press). In addition, from a social science standpoint the entire statement represents a kind of tailoring to the user group as a whole, because it addresses an issue that we know to be a special (but rarely expressed) concern of migraine sufferers. The inclusion of this text in the system is the outcome of a systematic attempt to understand more about the experience of migraine, and to present explanatory material that is sensitive to the perspective of patients. This example concerns a piece of knowledge about migraine that is well known to patients, but is not necessarily known to physicians. By addressing it in material generated by the system, we hope not only to meet an important information need of migraine patients, but also to help to empower them by communicating the fact that this concern is shared by others. As with all our explanatory material, the information sheet will contain a statement urging patients to raise significant concerns with their doctors.

6.2. Example 8 The next example concerns the opposite case: a piece of knowledge known to physicians but not necessarily to patients. Again, by making this knowledge explicit and bringing it to the attention of patients, we hope to support information exchange between doctor and patient. Physicians take for granted that they cannot predict with certainty how an individual will react to a given drug. This piece of medical common sense implies that doctors can't know for sure in advance which medication (or type of medication) will help a given migraine sufferer, nor can they know the dosage and dosage schedule at which a particular drug will be effective. Similarly, they cannot know for sure in advance whether patients will experience side effects from a given drug, nor do they know which side effects particular patients are likely to get. In short, the per-

414

D. E. Forsythe

spective of experts in neurology includes the expectation that successful treatment of migraine may involve a certain amount of trial and error. The fieldwork revealed that patients are not always aware of this fact. Like many nonphysicians, they often seem to believe that doctors k n o w - - o r ought to k n o w - - w h i c h drugs will help them. Similarly, unaware that the same drug may affect individuals quite differently, patients may expect physicians to know whether or not they will experience side effects from a given drug. Unless such faith in the omniscience of doctors is tempered, it surely contributes to the problem of compliance among migraine sufferers. Faced with a series of medications (or dosages) that have failed to alleviate their migraines, patients who assume omniscience may conclude that their doctor isn't trying, isn't competent, or simply doesn't believe that they suffer from a very painful condition. 5 (For illustrations of such reasoning, see Q10 and Q11 above.) In order to encourage patients not to draw such conclusions and drop out of treatment, we developed a piece of explanatory text designed to promote realistic expectations about the treatment process. This text will appear at the end of the interactive patient information sheet: Finding an effective treatment for migraine usually involves a process of trial and error. You and your doctor will need to work together to find the treatment that works for you. This may involve taking medications, making changes in your lifestyle, or both. Fortunately, there are many (treatments available for migraine), and the majority of patients are able to find one that is effective. The first medication the doctor prescribes may not be completely successful. Nevertheless, it is important for you to be patient and take each one as prescribed in order to give it a chance to work. Try not to get discouraged and don't give up: it is important to come back and try again until an effective treatment is found. This statement consists completely of canned text. As indicated by the angle brackets, it contains one mousesensitive area on which patients can click to see explanatory material concerning treatments available for migraine. 7. D I S C U S S I O N This article has attempted to demonstrate how ethnography (and an anthropological perspective) can be useful to system designers by helping them to understand the information needs and the perspective(s) of future users. It has illustrated how such ethnographic insights can be translated into explanatory material to be gener-

5Like others who suffer from chronic pain that has no visible cause, migraine sufferers are understandably sensitive to the possibility of disbelief. This problem is explored in Jackson (1992).

ated by a system. Using the example of the migraine project, it has also pointed to the reflexive role that anthropologists can play in a project. This final section comments on four themes in relation to this material: deriving design considerations from anthropological research, the utility (when appropriate) of changing assumptions in midstream, the importance of information exchange, and broadening (still further) our concept of explanation.

7.1. Deriving Design Considerations from Anthropological Research Reviewing this article, several computer scientists expressed reservations about applying conclusions from the study of h u m a n - h u m a n communication to the construction of systems for h u m a n - c o m p u t e r communication. They appeared to interpret our approach as implying that we equated the two types of communication. This reflects a misunderstanding. W e do not assume that h u m a n - c o m p u t e r communication is identical to communication between humans, nor do we see the latter as necessarily constituting a model for the former. On the contrary, one motivation for building the system in the first place is the realization that at present, doctor-patient communication does not always satisfy the information needs of migraine patients. The system can help to remedy this situation only if it provides explanation on topics of genuine concern to migraine sufferers. The primary goal of the ethnographic fieldwork is, therefore, to provide insight on what the system should explain; this requires that we identify the information needs of people with migraine. In order to present explanatory material in terms that make sense to users, we must also understand some broader aspects of their world view and experience. To accomplish these tasks, we have made use of a variety of data-gathering and interpretive m e t h o d s - - i n c l u d i n g observation of doctor-patient communication. As illustrated by Examples 1 - 3 in Section 3, the ethnographic data suggest inferences not only about what our system should explain to users, but also about how that material should be presented. We do not automatically draw design implications for the system from attributes of successful communication between humans. Rather, we view data on the latter as a source of ideas that may or may not work in the context of system design. For instance, Examples 1 - 3 suggest that it may be worthwhile to pay careful attention to tone and word selection in designing machine-generated explanatory material. As is clearly the case in h u m a n - h u m a n communication, it may be that users of explanation systems will be more easily enlisted if material generated by such systems addresses them in respectful and inclusive terms. Whether the use of such

Using Ethnography in Design language will actually affect the way users respond to our system remains to be seen, however. We intend to investigate this and other possible inferences for design drawn from the ethnographic data.

415 encounters the well-known problem of user acceptance. The ethnography has, therefore, enabled us to circumvent mistakes that would have been far more time-consuming and costly if discovered at a later stage.

7.2. Changing Assumptions in Midstream The type of rethinking that we undertook in the first year of the project is a predictable consequence of the systematic comparison between assumption and observable practice that constitutes an important part of ethnographic research. Experienced anthropologists expect that part way through a field project, they may discover they have been asking the wrong questions (Rosaldo, 1993, p. 7). Fortunately, the same analysis that leads to such a realization also helps to suggest what the right questions may be. However, since the use of ethnography is relatively new in system design, researchers in this field may be less accustomed to changing their major assumptions in midstream. During the first year of the migraine project, our systembuilding colleagues eagerly awaited the "results" of the ethnography. They expected these "results" to help them build the system they expected to build by providing data on how patients phrase their requests for textbook knowledge, as well as how best to word the textbook information offered by the prospective system. Instead, one of the first major results of the ethnography was to call into question our original problem formulation. ObservatiOn of what migraine patients actually do and say suggested that we had been thinking of building the wrong system. It is not easy to rethink basic assumptions and their implications for design. It required some adjustment to reconceptualize the proposed system as something other than a conventional tutoring system, that is, a system whose primary purpose i s t h e explication and transmission of formal, general information. Given what we now know about the information needs of migraine patients, however, it seems likely that the system we had planned to build would have encountered major problems of user acceptance. We were planning to build a system intended to target information needs that do not seem to be the most important ones from the standpoint of migraine patients. Fortunately, our observations have provided considerable information about what migraine patients do want to know. In other words, in addition to calling into ques tion some of our original design assumptions, the ethnography also provided some guidance in terms of which to redesign the prospective system. In addition, this rethinking was accomplished relatively early in the design process, at a time when we could easily reorganize our approach to system-building. This contrasts with the more usual situation inmedical informatics, in which inappropriate assumptions about users may not be discovered until a system is fielded and

7.3. The Importance of Information Exchange Research by Hunt et al. (1989b) suggests that noncompliance by patients relates less to misunderstanding on their part than to the fact that physicians and patients have different perspectives on issues beyond the purely medical. Since tacit assumptions of the sort described in this article help to constitute the perspectives of neurologists and their patients, these differences in what they take for granted may well affect compliance in migraine. If so, then compliance may well improve if information exchange between neurologists and migraine patients can be supported and expanded. This of course is one of the goals of our system. If we are successful in this attempt, then presumably more patient information needs will be met and a higher proportion of migraines will be successfully treated. We believe that machine-generated explanation can help to provide migraine patients with some of the information they seek including knowledge that tends to remain tacit during conversations between neurologists and patients. In addition to providing more conventional educational information about migraine treatments, triggers, medications, and side effects, the system we are building is intended to help bridge the gap between the perspectives of patients and neurologists by also making explicit part o f what each takes for granted. From a physician's standpoint, the patient encounter is an information-gathering session. TypiCally, the physician's perspective predominates. While the patient's perspective can provide a doctor with useful information, patients do not necessarily present what they know in terms that fit conventional medical categories. Understanding some of the assumptions that lie beneath physicians' questions and procedures should help patients to satisfy their doctors' information needs. At the same time, as we have seen. patients have information needs of their own that sometimes go unsatisfied. If an explanation system is truly to help support doctor-patient communication, it needs to support genuine information exchange between them. There are obvious limitations to computers' abilities to converse about an open-ended set of topics, such as those relevant to a patient's experience of chronic pain. Nevertheless, we believe that if a computer brings some acknowledgment of unexpressed information needs into an interaction with a patient, it may at least help to open up these topics for discussion with the physician thereby promoting more effective information exchange.

416

7.4. B r o a d e n i n g O u r Concept o f Explanation Knowledge-based systems are generally designed to meet information needs that are formal and general in nature (Forsythe et al., 1992). Less attention has been given to the problem of facilitating access to formal, specific information or to informal information that is either specific or general in nature. In the project described here, our initial approach reflected this convention: we simply took for granted that a patient education system in migraine should provide formal, general information about migraine and the drugs used to treat it. As w e have seen, however, data on the information needs of migraine patients have led us to realize that while migraineurs do need access to textbook-type information on migraine and its treatment options, they also need other information. Similarly, physicians need some information to which they do not routinely have access. All of this led us to consider whether and how our system could facilitate access by both patients and physicians to a much broader spectrum of information. In a previous study, we found that physicians often take for granted a narrow definition of "medical information,i' restricting the term to formal, general knowledge. Since observation o f information needs that actually arose in the course of patient care revealed a need for a wide range of information, including informal and/or specific knowledge that cannot be found in textbooks, we concluded that the definition of "medical information" should be broadened (Forsythe et al., 1992; Osheroff, Forsythe, Buchanan, Bankowitz, Blumenfeld, & Miller, 1991). Fieldwork in neurology for the present study has led us to a parallel realization with respect to the notion of explanation. The system under development for this project was initially conceptualized in terms of the current concept of explanation referred to in the Introduction (Section 1), which views explanation as a dialogue in which information is exchanged between the user and the system. However, using ethnography in the design of our explanation system has led to the realization that the concept of explanation needs to be broadened still further to include more types of knowledge in the dialogue. As this article has demonstrated, the relevant information to be exchanged in a medical setting includes not only facts about an individual patient and formal, general medical knowledge about disorders and treatments, but also some informal knowledge and some underlying assumptions that are not always expressed when doctors and patients communicate. If explanation systems are truly to support information exchange, they need to be able to accommodate these other types of knowledge as well.

A c k n o w l e d g m e n t s - - T h i s work is supported by grant number R01 LM05299 from the National Library of Medicine, National Institutes

D. E. Forsythe of Health. Its contents are solely the responsibility of the author and do not necessarily represent the official views of the National Library of Medicine. The fieldwork is carried out by D. E. Forsythe and M. Brostoff; analysis of texts derived from fieldwork sessions is carried out by D. E. Forsythe, M. Brostoff, and N. Bee. For helpful discus-

sions, I thank my colleagues on the research team: B. G. Buchanan (PI), M. Brostoff, N. Bee, J. Moore, V. Mittal, G. Carenini, S. Ohlsson, G. Banks, and S. Margolis. For helpful comments on earlier drafts of this manuscript, I thank D. Berry, B. G. Buchanan, V. Mittal, J. Nyce, B. Shen, and the anonymous reviewers.

REFERENCES Berg, M. (1987). Patient education and the physician-patient relationship. Journal of Family Practice, 24(2), 169-172. Buchanan, B. G., Moore, J., Forsythe, D. E., Carenini, G., Banks, G., & Ohlsson, S. (in press). Customizing information for interactive use by patients. Artificial Intelligence in Medicine. Buchanan, B. G., & Shortliffe, E. H. (Eds.) (1984). Rule-based expert systems. Reading, MA: Addison-Wesley. Carenini, G., & Moore, J. D. (1993a). Generating explanations in context. In W. D. Gray, W. E. Hefley, & D. Murray (Eds.), Proceedings of the International Workshop on Intelligent User Interfaces (pp. 175-182). Baltimore, MD: ACM Press. Carenini, G., & Moore, J. D. (1993b). Using the UMLS semantic network as a basis for constructing a terminological knowledge base: A preliminary report. In 17'h Symposium on Compater Applications in Medical Care (pp. 725-729). New York: McGrawHill, Fafchamps, D., Young, C. Y., & Tang, P. C. (1991). Modelling work practice: Input to the design of a physician workstation. In P. D. Clayton (Ed.), 15th Symposium on Computer Applications in Medical Care (pp. 788-792). New York: McGraw-Hill. Forsythe, D. E. (1992). Blaming the user in medical informatics: The cultural nature of scientific practice. Knowledge and Society, 9, 95-111. Forsythe, D. E. (1993a). The construction of work in artificial intelligence. Science, Technology and Human Values, 18(4), 460-479. Forsythe, D. E. (1993b). Engineering knowledge: The construction of knowledge in artificial intelligence. Social Studies of Science, 23(3), 445-477. Forsythe, D. E., & Buchanan, B. G. (1991). Broadening our approach to evaluating medical information systems. In P. D. Clayton (Ed.), 15'h Symposium on Computer Applications in Medical Care (pp. 8-12). New York: McGraw-Hill. Forsythe, D. E., Buchanan, B. G., Osheroff, J. A., & Miller, R. A. (1992). Expanding the concept of medical information: An observational study of physicians' information needs. Computers and Biomedical Research, 25(2), 181-200. Frankel, R. M. (1989). Talking in interviews: A dispreference for patient-initiated questions in physician-patient encounters. In G. Psathas (Ed.), Interaction competence (pp. 231-262) Lanham, MD: University Press of America. Geertz, C. (1973). The interpretation of cultures. New York: Basic Books. Geertz, C. (1983). Local knowledge: Further essays in interpretive anthropology. New York: Basic Books. Good, M.-J. (1992). Work as a haven from pain. In M.-J. Good, P. Brodwin, B. Good, & A. Kleinman (Eds.), Pain as human experience: An anthropological perspective (pp. 49-76). Berkeley: University of California Press. Hunt, L. M., Jordan, B., & Irwin, S. (1989a). Views of what's wrong: Diagnosis and patients' concepts of illness. Social Science and Medicine, 29(9), 945-956. Hunt, L. M., Jordan, B., Irwin, S., & Browner, C. H. (1989b). Compliance and the patient's perspective: Controlling symptoms in everyday life. Culture, Medicine and Psychiatry, 13, 315-334.

Using Ethnography in Design Jackson, J. (1992). "'After a while no one believes you": Real and unreal pain. In M.-J. Good, P. Brodwin, B. Good, & A. Kleinmann (Eds.), Pain as human experience: An anthropologicalperspective (pp. 138-168). Berkeley: University of California Press. Juge, C. F., & Assal, J.-P. (1992). Designing computer assisted instruction programs for diabetic patients: How can we make them really useful? In M. E. Frisse (Ed.), 16'h Symposium on Computer Applications in Medical Care (pp. 215-219). New York: McGraw-Hill. Kaplan, B. (1987). The influence of medical values and practices on medical computer applications. In J. Anderson & S. Jay (Eds.), Use and impact of computers in clinical medicine (pp. 39-50). New York: Springer-Verlag. Kleinman, A., Eisenberg, L., & Good, B. (1978). Culture, illness, and care. Clinical lessons from anthropologic and cross-cultural research. Annals of lnternal Medicine, 88(2), 251-258. Lane, P., McLellan, B., & Baggoley, C. (1989). Comparative efficacy of chlorpromazine and meperidine with dimenhydrinate in migraine headache. Annals of Emergency Medicine, 18, 360365. Ley, P. (1982). Satisfaction, compliance and communication. British Journal o f Clinical Psychology, 21, 241-254. Lundsgaarde, H. (1987). Evaluating medical expert systems. Social Science and Medicine, 24(10), 805-819. Mills, C. W. (1940). Situated actions and vocabularies of motive. American Sociological Review, 5, 904-913. Moore, J. D., & Ohlsson, S. (1992). Educating migraine patients: On-line generation of medical explanations in a migraine tutor. In Proceedings of the Fourteenth Annual Conference of the Cognitive Science Society (pp. 605-610). Hillsdale, NJ: Lawrence Erlbaum Associates. Moore, J. D., & Paris, C. L. (1992). Exploiting user feedback to compensate for the unreliability of user models. User Modeling and User-Adapted Interaction, 2(4), 331-365. Moore, J., & Swartout, W. (1989). A reactive approach to explanation. In Proceedings of the l lth International Joint Conference on Artificial Intelligence (pp. 1504-1510). Detroit, MI: Morgan Kaufmann. Moore, J., & Swartout, W. (1990). Pointing: A way toward explanation dialogue. In Proceedings of the 8 th National Conference on Artificial Intelligence (pp. 457-464). Cambridge, MA: MIT Press. Norman, D. A. (Ed.) (1993). Cognitive Science, 17(1), Special Issue on Situated Action. Nyce, L, & Bader, G. (1993). Hierarchy, individualism and hypermedia in two American high schools. In L. Ingelstam & L.

417 Sturesson (Eds.), Brus over landet. Om informations over flodet, Kunskapen och manniskan (pp. 247-259). Stockholm: Carlsson. Nyce, J., & Graves, W. (1990). The construction of knowledge in neurology: Implications for hypermedia system development. Artificial Intelligence in Medicine, 2, 315-322. Nyce, J., & Timpka, T. (1993). Work, knowledge and argument in specialist consultations: Incorporating tacit knowledge into system design and development. Medical and Biological Engineering and Computing, 31(January), HTAI6-HTAI9. Osheroff, J. A., Forsythe, D. E., Buchanan, B. G., Bankowitz, R. A., Blumenfeld, B. H., & Miller, R. A. (1991). Physicians' information needs: Analysis of clinical questions posed during patient care activity. Annals of Internal Medicine, 14(7), 576581. Rosaldo, R. (1993). Culture and truth: The remaking of social analysis. Boston: Beacon Press. Saper, J. R., Silberstein, S. D., Gordon, C. D., & Hamel, R. L. (1993). Handbook of headache management: A practical guide to diagnosis and treatment of head, neck, and facial pain. Baltimore, MD: Williams and Wilkins. Stang, P., & Osterhaus, J. (1993). Impact of migraine in the United States: Data from the National Health Interview Survey. Headache, 33(1), 29-35. Suchman, L. (1987). Plans and situated actions. Cambridge, England: Cambridge University Press. Suchman, L. (1990). Representing practice in cognitive science. In M. Lynch & S. Woolgar (Eds.), Representation in scientificpractice (pp. 301-321). Cambridge, MA: The MIT Press. Swartout, W. R., & Moore, J. D. (1993). Explanation in second generation expert systems. In J.-M. David, J.-P. Krivine, & R. Simmons (Eds.), Second generation expert systems (pp. 543585). New York: Springer-Verlag. Tannen, D. (1986). That's not what I meant.t: How conversational style makes or breaks your relations with others. New York: Morrow. Traweek, S. (1993). An introduction to cultural and social studies of sciences and technologies. Culture, Medicine and Psychiatry, 17, 3-25. Wallen, J., Waitzkin, H., & Stoeckle, J. D. (1979). Physician stereotypes about female health and illness. Women Health, 4, 371388. Werner, O., & Schoepfle, G. (1987). Systematicfieldwork. Newbury Park, CA: Sage. West, C. (1984). Medical misfires: Mishearings, misgivings, and misunderstandings in physician-patient dialogues. Discourse Processes, 7, 107-134. Wynn, E. (1991). Taking practice seriously. In J. Greenbaum & J. Kyng (Eds.), Design at work (pp. 45-64). Hillsdale, NJ: Erlbaum.