Bundles in the Wild: Managing Information to Solve Problems and Maintain Situation Awareness.

We cannot know what the task is until we know what the tools are. (Hutchins, 1995, p. 114)

ABSTRACT

THIS ARTICLE DESCRIBES HOW EXPERTS CREATE and use bundles--organized, highly selective collections of information--to help solve problems and maintain situation awareness. In field observations of expert clinicians caring for patients in critical care units, bundles appear to be a widely used means of managing information to support diverse, complex, and often simultaneous tasks. They may be especially useful in settings that are characterized by high uncertainty, low predictability, frequent interruptions, and potentially grave outcomes; where time and attention are highly constrained; and where interdisciplinary teamwork is essential. Reports of analogous observations from other domains such as aviation and air traffic control suggest that bundles may be a common information management tool for solving problems and maintaining situation awareness. In an age of digital libraries, computer-based tools for creating and managing bundles may be useful as the information in these settings is increasingly represented in digital collections that are larger, more complex, more diverse, and potentially more difficult to explore and manipulate.

BACKGROUND

Digital Libraries Research: Tracking Footprints in Information Space

The observations described in this article are part of a larger project, "Tracking Footprints in a Medical Information Space: Computer Scientist-Physician Collaborative Study of Expert Problem Solvers" funded by the National Science Foundation Digital Libraries Initiative Phase 2 (National Science Foundation, 1998). The goals of the Tracking Footprints project are to understand how experts select information in a large and complex information space and to develop tools that assist them in this process. The research focuses on experts in health care, but analogous observations by others suggest that the findings will be applicable in other domains such as aviation. The research is being conducted by two teams--an observation team whose job is to accurately describe the information behavior of expert clinicians in situ, and a computer science team whose job is to investigate the application of superimposed information technology (Delcambre & Maier, 1999) to assist experts in navigating vast and complex digital information spaces. It is important to explicitly state the underlying assumptions of this research and to distinguish it from other areas of digital library research and other uses of digital libraries. Most important here is to distinguish the focused information seeking of clinical problem solving from other information behaviors such as browsing or information gathering (Krikelas, 1983) that may be observed in other uses of digital libraries.

Assumptions: Framing the Problem

Imagine a heart specialist who is called in to see a patient to manage a specific heart condition. While reviewing the medical records of the patient, the specialist must somehow locate sufficient relevant information to understand and solve the problem, ignoring the much larger quantity of information that belongs in the record but is irrelevant or redundant with respect to the problem at hand. As she traverses this large, diverse, often disorganized collection of documents, she makes explicit choices about which items to ignore and which to examine more carefully. Taken together, her choice's create a discrete subset of information and documents that are relevant to a given problem and likely to be of interest to other users of the collection who are concerned with the same problem.

The "user" in this case is an expert or team of experts, possessed of specialized knowledge, focused on a specialized patient care task. Information management, although essential, is of secondary importance compared to the clinical task. Significant constraints are present--i.e., time and attention are quite limited; considerable uncertainty and unpredictability are present; and misunderstanding or error have potentially grave consequences. To be effective under these constraints, expert clinicians must employ cognitive strategies such as hypothetico-deductive reasoning (Elstein, Shulman, & Sprafka, 1978) to narrow the problem space, and "satisficing" (Simon, 1955) to find a solution that is satisfactory rather than devote substantially more time and attention to arrive at an optimal one. It is expected that only a small fraction of the available information will be examined, the vast majority of the collection will be ignored, and these choices will often be based on the appearance and organization of the documents rather than on their content (Nygren & Henriksson, 1992; Nygren, Lind, Johnson, & Sandblad, 1992).

The task, involving some aspect of patient care, is likely to be both generic and unique. It is assumed that most users in most circumstances need similar information to make a certain type of decision or to perform a certain task. But in health care especially, every instance is expected to have unique elements, relating to the patient, to the clinician, or to local circumstances, and this variability cannot be completely predicted from prior instances. As Sir William Osier put it nearly 100 years ago: "Variability is the law of life, and as no two faces are the same, so no two bodies are alike, and no two individuals react alike and behave alike under the abnormal conditions which we know as disease" (Osier, 1932).

The information space for this task includes the medical record of the patient (see Figure 1)--a large complex collection of information distributed across multiple, often geographically dispersed, information systems--some electronic, some paper--created by a diverse array of health professionals, for divergent purposes, over an extended period of time. Different classes of users may inhabit distinct territories within this space, rarely venturing into other regions (Ames, 1993). Importantly, the medical record itself is only one of many sources of information, electronic and otherwise, that are likely to be employed in the management of clinical problems (Gorman, 1999). Sorting through these sources to locate needed information can be a formidable and time-consuming task. As electronic health information systems, like digital libraries, expand and evolve, their diversity, size, and complexity are increasing dramatically while familiar cues that enable efficient navigation in print media disappear (Nygren & Henriksson, 1992; Nygren et al., 1992). The result may be an information management task that is even more formidable than before.

[Figure 1 ILLUSTRATION OMITTED]

Research Questions

A common observation in our early pilot work for this project was that clinicians in a variety of settings select and organize bits of information into what we are calling bundles (Figure 2). As a result, we refined our research questions to focus on the following:

1. How do experts choose which items to examine and which to ignore?

2. (a) Is there value in the selection and organization of information items into bundles?

(b) Can a bundle be reused by the expert who created it?

(c) Can a bundle be used by other experts?

3. Can we capture and leverage the information inherent in bundles?

[Figure 2 ILLUSTRATION OMITTED]

Initially, we considered devising protocols to examine questions 2a, b, and c in a laboratory environment. However, our observations of expert clinicians in subsequent fieldwork has provided the opportunity to answer these questions "in the wild" (Hutchins, 1995).

INFORMATION USE IN CRITICAL CARE

This study employs observational methods to address the general question, "How do expert clinicians use information to help patients in critical care?" Complete details of the methodology of this study will be available in a forthcoming report. In this discussion, we briefly review the methods and describe in detail those findings related to the creation and use of bundles by experts as they care for patients in an intensive care unit.

Setting and Subjects

This study was conducted in the intensive care unit (ICU) and the cardiac intensive care unit (CICU) of a university-affiliated metropolitan hospital of moderate size. This institution serves as a regional referral center for smaller community hospitals and as a training institution for multiple professions, including medical and nursing undergraduate students, an internal medicine residency training program, and a variety of postgraduate and advanced practice programs in nursing and medicine. The ICU and CICU employ a common layout with individual patient rooms on the perimeter of a rectangle surrounding a central work zone for nurses, physicians, and others. Each patient's room has its entrance and a transparent wall facing the central area maximizing the ability of clinicians to directly observe the occupants and equipment of every room from the central area (views of the setting and many observed artifacts can be found on the project Web site at: http://www.cse.ogi.edu/footprints/).

This clinical setting is characterized by a high level of patient complexity and acuity; a considerable and unpredictable flow of patients in and out of the unit; complex medical equipment of every description to support a variety of patient care and other tasks; a constant stream of diverse hospital personnel and visitors from outside the unit; a remarkably high level of ambient noise; and a professional team approach that is highly focused on patient care in a setting of constant change, interruption, and uncertainty.

Participants in the study included the critical care nurses, resident physicians, attending physicians, pharmacists, therapists, and an array of other health professionals that are typically engaged in modern critical care. Consent was obtained from informants in group meetings at the outset of the study. By coordinating with unit leaders, observation sessions were scheduled only when staff who were willing to participate were on duty. Individual staff or patient identifying data either were not collected or have been obscured where they are inadvertently included in images, and so on.

Data Collection

Data were collected mainly through participant observation, supplemented by key informant interviews, oral histories, and focus groups. Data included field notes, taped transcripts of interviews, photographs, and video recordings. The data collection team included a nurse (ML), a medical librarian (JA), a physician-in-training (JL), and an attending physician (PG). The team leader (PG) is a physician on the attending staff of the host institution, enabling the team to "gain entry" to the unit more readily. The mix of professions on the team afforded the ability to match observers and participants by professional background, promoting better rapport with informants and more valid observations. This mix of professions also ensured that multiple perspectives would be represented in the examination of data, improving the reliability of our analysis. Field observations and data analysis were guided by the general research question, "How do expert clinicians use information to help patients in critical care?" Following analysis by the research team, findings were presented to informants in small groups to obtain a validating "member check" on our observations and inferences.

OBSERVATIONS: BUNDLES IN THE WILD

A full report of the findings of this study is forthcoming. We focus here on observations relating to the creation and use of bundles. Following are descriptions of several types of bundles which expert clinicians used in the course of patient care.

The Kardex

The Kardex (Figure 3) captures the current state of active medications, treatments, and other management information for a given patient. One Kardex is created and maintained for the duration of each patient's stay in the ICU, then discarded on transfer to another unit. Printed on heavy stock and folded, it is durable, portable, and separate from other documents, enabling it to survive frequent revision and heavy use as it is taken along to the bedside, to the workstation, to the pharmacy, or to the telephone in support of various tasks. Identifying information is stamped in ink using an addressograph, as it is on all patient-specific documents. Virtually all other entries are made in pencil. This permits inactive medication or treatment orders to be erased and new information to be added so that only current active information is present at any given moment. The primary nurse caring for each patient is responsible, with help from the unit secretary, for keeping that patient's Kardex up-to-date, although many others, including physicians, nurses, pharmacists, etc., may refer to the Kardex in the course of carrying out their respective duties.

[Figure 3 ILLUSTRATION OMITTED]

Most of the information recorded on the Kardex is also recorded in another location, such as in the nursing care plans, in the physicians' orders, in pharmacy records, in laboratory records, and so on. But no other source contains this particular combination of information. Some information is recorded only on the Kardex. This information tends to be details that are helpful to nurses and others caring for the patient but which are too sensitive, uncertain, unique, or temporary to be documented in the permanent record. While copies of the information on the Kardex exist in other locations, only one Kardex exists for each patient. This provides a physical form of version control: there is no confusion about whether the Kardex in hand is the current and correct version, a problem that can occur with other documents of which multiple versions may exist and multiple copies may be printed.

The layout of the Kardex provides predictability in the location of information but allows flexibility of content according to the unique and changing patient situation. Frequent revisions are made to keep the Kardex current, complete, and relevant. Observation reveals that it is used frequently by nurses, physicians, and other personnel. In the ICU under study, the Kardex was reported to be among the most useful sources of information by residents (unpublished data). As one nurse observed, "if you talk to people, it's probably their number one communication tool." In many institutions, efforts are being made to eliminate the use of the Kardex (Krause, Westdorp, Coonen, & Jenks, 1996; Newell, 1996).

A Resident's Worksheet

Resident physicians often create an informal temporary daily worksheet containing selected task-oriented information (Figure 4). Although styles vary, these generally are preprinted or photocopied forms, often using a format borrowed or modified from that of a colleague. In the example in Figure 4, each row contains information about a single patient and each column contains information of a certain type: in the leftmost column is identifying information, often only the last name and room number; in the second column, a list of active health problems; in the third column, pertinent, often numeric, laboratory data; in the rightmost column, action or to-do items. In each cell, only a subset of the available information is included: complete information can be found elsewhere. For example, the active problem list cell usually includes only those conditions that require action or attention while in the ICU. A more complete problem list may be found in the medical record but is not necessary on this worksheet.

[Figure 4 ILLUSTRATION OMITTED]

The figure illustrates annotations that suggest the meaning of unlabeled numeric or text data. Some of these annotations, such as the simple matrices used to record standard hematologic and metabolic parameters, are widely used and readily understood by other clinicians. Other annotations may be idiosyncratic or ad hoc and understood only by their author or in context. The content of these worksheets is generally unique to the individual patients, to the individual resident, and to the time it is created: what is relevant and important to one clinician on one shift may not be relevant and important to another clinician or at another time. Often there is a mix of patient-specific data and informal or local procedural information (Forsythe, Buchanan, Osheroff, & Miller, 1992; Gorman, 1995).

There is an emphasis on recording numeric laboratory data, perhaps because these numeric details are less easily retrieved from memory. In contrast, the simple list of abbreviations or phrases in the "Active Problems" column can trigger retrieval of additional details from memory, either information obtained by direct experience (hearing the history from the patient, observing physical findings through physical examination) or information that is more amenable to "chunking" into high level "illness scripts"' (Evans & Gadd, 1989; Schmidt, Norman, & Boshuizen, 1990). Graphical cues, annotation, and white space are used: (1) to specify item types (check boxes for to-do items); (2) to specify data types (the shape of simple matrices identify the type of laboratory data), and (3) to draw attention to important or needed information (empty matrices suggest a need to retrieve specific laboratory data). Although many elements are recorded for all patients, the combination of data recorded is unique to the patient and the situation. Placeholders are not present for data that is not expected to be needed.

This bundle is like the Kardex in several respects: (1) it is a representation of the current state (active problems and current laboratory data); (2) it serves to organize the performance of important tasks; (3) its structure has predictability but great flexibility for information content; (4) most of the data have been or will be recorded elsewhere; (5) the data are highly selected, with more complete details available in other documents; and (6) there may be a mix of patient-specific data (found in the print or electronic patient record) and local or informal procedural information (often communicated orally or found in local information sources). Also like the Kardex, it is a portable, temporary, and highly task-oriented document. Unlike the Kardex, it contains information about multiple patients, it is usually created and used by a single individual, and is not generally shared. Similar resident worksheets were observed which were shared, but these often contained only baseline information with unique annotations added by individual recipients of partially completed bundles.

Messy Bundles

While both the Kardex and resident physician's worksheet are prepared in advance, with some a priori sense of what to record and where and why to record it, other bundles are ad hoc, literally "back of the envelope" creations (Figure 5). We refer to these as messy bundles, with the messiness occurring on multiple levels: from the informality of the recording medium, such as this "four by four" gauze pad package, to the shorthand used to record it, often without labels or formatting; and even perhaps to "messiness" or uncertainty about these data. Messy bundles may be created by anyone and reused by their author or by others, but they tend to be very tightly integrated with tasks, and their interpretation depends much more on context than is true of other types of bundles.

[Figure 5 ILLUSTRATION OMITTED]

Messy bundles may be advantageous beyond mere convenience because of their immediacy, portability, disposability, and flexibility. Potential advantages of messy bundles may include: (1) where there is uncertainty about the data, it can be verified before it is added to the permanent record; (2) the act of physically recording the data may improve understanding, allowing the clinician to "think through" the information as it is being recorded; (3) messy bundles can be taken to the bedside, to the medication room, wherever the information is needed, to whomever needs it, without interrupting patient care; (4) messy bundles can be kept handy as a physical reminder of patient status, of the need to perform some task, or that the task has been performed; (5) messy bundles allow information to be recorded in whatever form is appropriate without the limitations of the recording technology, which may be especially useful for such items as waveforms, diagrams of physical findings, or the location of equipment in relation to the patient; and (6) assembly of messy bundles can be readily integrated into routine workflow, adding items of information as they become available at separate locations, often distant from formal documentation technology.

Another advantage of messy bundles is that their flexibility allows for a very high signal-to-noise ratio, a concept borrowed from information theory referring to the amount of meaningful information ("signal") conveyed in a message relative to the amount of useless information ("noise") in that message. The bundle in the figure contains seventeen numeric values, only two of which are labeled. Yet, in informal feedback sessions, physicians and ICU nurses readily recognized these to be hemodynamic data from a patient with a pulmonary artery catheter. In a critical care unit, the terms "CVP" and "Wedge" have specific meanings that help provide context for the other data. For the unlabeled items, the range of values, the order in which they are recorded, the use of separator characters and annotation (encircled groups, bars over "means"), and the grouping of items provide the necessary cues for the data types to be understood by experts in this domain, even if undecipherable to others. Evans and Gadd (1989), in their analysis of expert discourse, comment on this sort of efficiency in expert communication: "Experts share knowledge not only of the details of their domains, but also of the structure and goals of their discourses" (p. 214).

The ICU Flow Sheet

The flow sheet (Figure 6) is a nearly ubiquitous tool for managing information in intensive care units. Unlike the first three examples, a flow sheet is both an essential working document and a part of the permanent record. Compared to the other bundles, it is less portable, not disposable, more structured, and less flexible in its content and organization, although much flexibility and some portability remain. It is a dynamic document, evolving over the course of a shift or a day as it is detailed with data about the patient. Like the Kardex, only the nurse primarily responsible for the patient makes entries on the flow sheet, while nearly every clinician involved in caring for the patient makes use of it.

[Figure 6 ILLUSTRATION OMITTED]

The data content is highly predictable but still sufficiently flexible to remain context-specific and serve the needs of the individual patient and clinicians. For example, vital signs are always recorded, but the frequency and means of recording them depends on the state of the patient. Medications, various nursing observations, hemodynamic measurements, or laboratory data may or may not be included as dictated by the current context of the patient's care. Highly abnormal values, rapidly changing parameters, or measurements used to monitor an active treatment or condition will be included, but only for as long as is necessary. Experts using the flow sheet can determine at a glance the overall severity and stability of the patient's condition as well as the major conditions that require attention at a given point in time.

The data format is highly structured, using a row and column arrangement. Data in each column roughly represent a cross-section in time, with time moving from left to right across the page. The time scale may be expanded or contracted according to the frequency of data collection required by the patient's condition. Rows are grouped to contain related data: temperature, pulse, and blood pressure at the top, nursing observations and physiologic parameters beneath these, medication information and laboratory data in groups of rows near the bottom, and text annotations added wherever meaningful. Common symbols, figures, abbreviations, annotations, separators, and data arrangements are used. There is general predictability but specific flexibility in the recording of data items. A domain expert from another institution can readily interpret flow sheet information with no training and minimal explanation, partly because of expertise about the data, partly because of experience with similar documents used in other institutions.

Nearly every individual item of information is recorded in other documents or information systems, but the flowsheet contains an assembly of information collected asynchronously in the course of workflow from diverse, physically separate, sources, and the flowsheet contains alternate representations of the data that are meaningful and useful in task-specific ways, using graphical plots, symbols, diagrams, numbers, text annotation, and meaningful arrangements or juxtaposition. The current dose of powerful medications is recorded alongside current physiologic measurements that influence or are influenced by that medication, allowing minute-to-minute adjustment. Text annotations such as "questionable waveform," "vomiting, possible aspiration," or "family visiting" are recorded immediately adjacent to numeric or other data, allowing those data to be understood in a way that would not be possible otherwise. Even juxtaposition in time of items that may not otherwise appear to be related (and therefore would not be recorded together) may enable patterns to be recognized that could otherwise go unnoticed. Like the Kardex, many institutions are replacing the flow sheet with electronic systems that attempt to emulate the content, format, and/or functions of the flowsheet (Shabot, 1997).

A Bundle of Bundles

Groups of bundles may be positioned together in temporary arrangements, creating what may be referred to as a metabundle as illustrated in Figure 7. Bundles may be present, such as the large flowsheet and Kardex, alongside other nonselective information collections, such as the patient's chart (three-ring binder) and electronic information systems in varying combinations. The entire workstation is immediately adjacent to the patient's room so that the patient and all monitoring and treatment devices are in view at all times from this location. Once again, this is primarily a workstation for the critical care nurse, but many other expert clinicians use it: physicians, pharmacists, therapists, and others. Unwritten socially negotiated rules of "ownership" allow for some objects to be shared or removed for use elsewhere, while others are left in place except under unusual circumstances. It is rare, for example, to observe a physician recording information on a flow sheet, but it is common to see physicians and others reading from it. Redundant recording of information provides a degree of fault-tolerance, allowing certain items to be removed for task-specific purposes while keeping essential data available at the bedside.

[Figure 7 ILLUSTRATION OMITTED]

The content and arrangement of this metabundle is quite flexible and meaningful. What is lost in predictability of content is gained in meaning: the presence and arrangement of objects contains information about the present state and immediate past and future states of the patient. Implicit in the positioning of information objects is a to-do list and a priority classification. Lab slips pinned to the doorframe suggest the need to collect specimens. Hemodynamic data left out in full view suggest that: (1) there is a reason for this patient to have it recorded; (2) the task of obtaining it has been completed; (3) the task of interpreting it is not yet complete; and (4) the attention, of an appropriate expert may be needed. Laboratory report slips positioned neatly out of view beneath Kardex on the clipboard suggest that these data have been obtained, recorded, examined, and probably interpreted to some extent. If a major important abnormality were present, the report would likely be on top of the clipboard; if the data were normal, stable, and unlikely to require further attention, the paper would likely be filed or recycled. The computer screen, showing a clinical information application rather than a screen saver pattern, suggests recent use.

To describe this as a metabundle is to focus only on the information objects and information systems that are present. But to do this is to create an artifact, artificially separating the information objects from the people and the other objects that are integral to the tasks. The location of this metabundle is essential: its usefulness is in large part a function of where it is, and to move it could significantly alter its ability to support the tasks of patient care. Similarly, the other objects that are present are not traditionally viewed as information objects, but they are integral to performance of patient care tasks, and their presence and physical state (open versus closed, and so on) contain information about the state of the patient or patient care tasks. Airline pilots use the same sorts of cues; for example, the physical position of controls in the cockpit can be a representation of the current state of the object being controlled and may have implications for the current stage in completion of a task (Hutchins & Klausen, 1996).

Not visible in the figure are the sounds--i.e., regular constant sounds that suggest stability; irregular or intermittent sounds that suggest change and draw attention; and alarms that indicate a problem and demand attention. Any of these, by virtue of their volume and directionality, provide a direct analog representation of the problem: its location, its importance, and its relevance to the listener. Though often not viewed as elements of an information system, these are important sources of information that can be essential to effective and efficient performance of critical tasks.

SUMMARY OF BUNDLE PROPERTIES

The goal of this article is to describe bundles--organized collections of highly selected information--which we observed to be commonly used by expert clinicians to help solve problems and maintain situation awareness. Based on fieldwork in an intensive care unit, five examples of bundles have been described that illustrate the variety of bundles in use; their information content; organizational format; common and distinguishing features; and usefulness in the intensive care environment for solving problems; performing patient care tasks; and developing, maintaining, or sharing situation awareness. To clarify what bundles are and what they are not, the properties of bundles can be summarized as follows:

Bundles are always:

actively created   A bundle is by definition actively created by an
                   expert in the course of problem solving, task
                   performance, or maintaining situation awareness.
physical           Some important information exists only as an
                   individual or collective mental representation
                   but, for our purposes, a bundle is always a
                   physical representation.
collections        Creating a bundle involves grouping items, based
                   on some idea of relatedness. A single item would
                   be trivial and require no processing.
selective          Creating a bundle involves selection, separating
                   relevant and important from nonrelevant or
                   unimportant. A complete and comprehensive
                   collection would be of far less value, possibly
                   interfering with task performance.
multigranular      Most bundles are comprised of bits and pieces of
                   data of varying levels of granularity from
                   single items, to groups of contiguous or related
                   items, to a paragraph or more of information.
                   Rarely does a bundle incorporate a complete
                   document.
context specific   Bundles often share common data elements or
                   format, but every bundle is to some extent
                   context dependent, where context may include
                   properties specific to its creator, the
                   setting, the patient and task, and a specific
                   point in time.
                   Bundles are therefore never completely
                   predictable and must be highly flexible.
task oriented      Bundles are created not for their own sake but
                   to facilitate performance of specific tasks.
                   Creating a bundle may: (1) provide a data cache
                   for specific items especially where precision
                   is required as with laboratory values physiologic
                   parameters;
                   (2) trigger memory of additional details;
                   (3) facilitate development of an individual's
                   situation awareness (getting the story straight,
                   thinking the problem through); (4) enable
                   sharing of related data with others having
                   related expertise and duties, achieving group
                   situation awareness; and (5) facilitate
                   multi-tasking and multi-threading in an
                   environment characterized by simultaneous demands
                   and frequent interruptions.
redundant          Almost every item of information in a bundle can
                   be found elsewhere: in another bundle; in other
                   documents, collections, or information systems;
                   or in some physical representation. The bundle
                   itself is a unique collection that can be found
                   nowhere else.
Bundles are
sometimes:
organized          All bundles probably have some degree of
                   organization, distinguishing them from
                   unordered collections. Bundles observed in this
                   study had varying degrees of organization:
                   (1) high degree of order in a fixed, widely
                   recognized pattern; (2) high degree of
                   organization in an individual or locally unique
                   order; and (3) loosely organized with fuzzy or
                   uncertain relationships. The order in bundles is
                   often latent, emerging as the bundle is created.
dynamic            Some bundles are static: created and used, then
                   stored or discarded. Others are frequently
                   revised (the Kardex) or regularly updated (the
                   flow sheet) as the situation changes.
temporary          Some bundles are for temporary use, others
                   become a part of the permanent record with
                   implications for the kinds of data they can
                   contain.
diverse            Most bundles include information items derived
                   from multiple sources or systems, although many
                   are single-source.
complex            Many bundles contain multiple data types or
                   multiple media. Numeric data, text, annotations,
                   diagrams, cutouts or replicas of waveforms, and
                   a variety of graphic elements may be included.
multiauthored      Many examples of composite, multiauthored
                   bundles exist, though those illustrated here
                   tend to be single author.
shared             Some bundles are created and used by a single
                   person, but many are used by multiple
                   individuals, often crossing a variety of
                   disciplines.
multipurpose       Bundles may serve one or many tasks.
uncertain          Some data or the suggested relationships among
                   data are tentative or uncertain. Bundles provide
                   a means of dealing with this uncertainty by
                   making it temporary or keeping it private, for
                   example.
visual             Almost all bundles we observed made use of
                   visual cues (layout, ink color, handwriting
                   style, white space, check boxes, arrows,
                   circles, etc.) to organize information, draw
                   attention to specific items, suggest
                   relatedness or patterns, and otherwise
                   provide a means of adding meaning and focusing
                   the attention of viewers. These annotations
                   tended to be graphical, flexible, and
                   were sometimes novel.

Bundles are not:
distributed        Bundles by definition are items brought together
                   in one place, although the items may be gathered
                   from widely distributed sources.
comprehensive      As stated above, exhaustive lists, though
                   important for archival and other purposes,
                   appear to be of less value than selected
                   task-specific information.
simple lists       Bundles have some internal organization,
                   determined by the context in which they are
                   created and by the ontology of their creator's
                   domain. This internal organization thus cannot
                   be completely predicted in advance.
forms or views     Views or forms can provide highly meaningful
                   grouping of information items but, unlike
                   bundles, are less context specific and are
                   entirely predictable in their content and
                   arrangement.

BUNDLES AND DIGITAL LIBRARIES

As libraries evolve from physical repositories in fixed locations to digital collections available in virtually any location, it can be expected that they will be used in new ways beyond traditional information gathering or research. In particular, because digital libraries can provide vast amounts of information at the times and places where information intensive tasks are performed, they have the potential to transform the work of those whose information management and decision making are tightly integrated with the physical operations they perform. To achieve a truly beneficial transformation, however, requires a deep understanding of the work, the experts who perform it, and the information tools and processes that they use. As Ruhleder (1994) states: "Medium, thought, artifacts, and work processes are deeply intertwined, often in ways we do not fully understand" (p. 210).

Our observations of expert clinicians using bundles to manage information while caring for patients in an ICU are analogous to observations from other fields. Levy and Marshall (1995), for example, describe how information analysts in a digital library "create more fluid, transient, and nondigital materials, constructing and maintaining local collections, which can then be shared with others" (p. 80). These observations may provide insights into some of the challenges that digital libraries must address if they are to become truly useful in their new roles. Five features of bundles that appear to be especially important in this regard are: tangibility, informality, redundancy, annotation, and active creation.

Tangibility refers to the physical properties of bundles and the task-specific advantages these physical properties confer. As portable objects, they can be kept on-hand at-hand and thereby allowing information to be collected, referred to, and shared as individuals move about to perform tasks that are often exacting, messy, or physically demanding. For groups collaborating on multiple simultaneous complex tasks, they provide a tangible means of version control (the Kardex); socially negotiated authorial control (as with the flow sheet); a visible means of determining and transferring responsibility both for the bundle and for the task it is created to support; and a physical representation of the state of a task or tasks that enables multi-threading in the face of frequent interruptions, reshuffling of priorities, and reassignment of responsibilities. Similar observations have been reported in other domains, notably among groups of air traffic controllers who rely on paper flight strips, "physical objects with multiple affordances that support various aspects of the controllers' work" (Mackay, 1999, p. 323).

Informality refers to the flexibility, on multiple levels, that bundles provide their users, including flexibility of content, format, granularity, ownership, and so on. As temporary disposable collections, they can contain information that does not belong in the permanent record: sensitive information that is meant to be kept private; logistic information with only short-term contextual usefulness or relevance; and combinations of information types (Gorman, 1999) that must be integrated to perform clinical tasks but must be stored separately in formal information systems. As informal creations, bundles allow individuals or groups to manipulate information in tentative combinations, avoiding the substantial cognitive overhead demanded by computer systems that require formal specification of categories and relationships (Shipman & Marshall, 1999). The "pile" metaphor for a file system interface (Rose et al., 1993) and computer-supported incremental formalization (Shipman & McCall, 1999) are two examples of information systems designed to allow for, and benefit from, informal specification by users.

Redundancy refers to the use of multiple representations of data and multiple processes or procedures for managing it. Most of the data present in bundles are also present in at least one other information system. Many of these items are collected, recorded, or shared by more than one person or through more than one process. One advantage to this use of multiple representations and multiple processes is fault tolerance: bundles are one tool used to check and double check information that may be questionable, unavailable, or lost. Experts thus take advantage of one another's expertise and attention to make sense of the information and make certain it is correct (Figure 8). Similar fault tolerance has been described when a group receives communication in an airline cockpit (Hutchins & Klausen, 1996) and when the presence of aircraft is simultaneously apparent on a radar screen and as a physical flight strip (Mackay, 1999).

[Figure 8 ILLUSTRATION OMITTED]

A second benefit to this use of multiple representations is the task suitability of the representation. As Shipman and Marshall (1999) note, "a representation that is suitable for one task may not be appropriate for a very similar related task" (p. 345). Hutchins describes this phenomenon on the navigation deck of an aircraft carrier. Bearings and ranges are simultaneously recorded in numeric tabular form on the ship's log for documentation purposes and separately plotted as a direct analog representation on the navigation chart, a form that is more immediately useful for navigation purposes (Hutchins, 1995). This process of "propagating of representational state across a series of representational media" (Hutchins, 1995) appears to be analogous to some of the bundles created in the ICU, where bundles allow multiple representations, each suited to a specific task.

Annotation here refers to the ways in which experts add meaning to a collection of information. The importance of annotation and the usefulness of a technology (currently paper) that enables it has been described by others. Levy and Marshall (1995) report that: "In spite of organizational efforts to make all sources available through digital means ..., analysts still make extensive use of paper as the principal interpretive medium" (p. 81). Three forms of annotation appear important here: selection, organization, and the addition of symbols. Selection alone can be considered a form of annotation, adding value through data reduction, where expert knowledge of what is relevant and important is inherent in the selection itself. Organization of selected information is another form of annotation, adding meaning by positioning items in a manner that suggests relationships or categories. This may occur within bundles, where the location of items may impart meaning (Figures 4 and 5) or between bundles, where the location of documents suggests meaning (Figures 6 and 7). Experts in other domains use spatial positioning in this way, including air traffic controllers (Mackay, 1999) and information analysts, who "express nuances of meaning by simply juxtaposing paper documents" (Levy & Marshall, 1995, p. 81). Shipman and Marshall (1995) have developed a "spatial parser" to infer meaning inherent in spatial organization. The third and most obvious form of annotation is the addition of symbols, including text, evident in the bundles described in this discussion (Figures 4 and 5). Marshall (2000) has provided a detailed discussion of the challenges to digital libraries that annotation presents.

Active creation is a final property of bundles that appears to be important to their usefulness. Each of the bundles described in this article was actively created to support one or more specific tasks. In focus groups, informants in this study commented on the benefit of this activity, allowing active processing of information, improving their understanding and awareness. In related work, physician trainees commented that the process of organizing patient information into a single "History and Physical" document at the time of hospital admission enhanced their understanding of the case, and this enhanced understanding was lost when responsibility for creating this document was transferred to others (Ash, Gorman, Hersh, Lavelle, & Poulsen, 1999). A related phenomenon has been noted in aviation automation, where it has been suggested that cockpit automation, such as Ground Proximity Warning Systems (GPWS), by removing humans from the information processing loop, may diminish situation awareness by the flight crew and thereby fail to achieve the intended goal of reducing the danger of Controlled Flight Into Terrain (CFIT) (von Thaden, 2000). It is thus an open question whether automatically generated bundles, with information selected, organized, and annotated without human participation, would be as useful as the real thing.

BUNDLES IN CAPTIVITY: SLIMPAD

Based in part on these observations, the computer science team in our group has been exploring the issues associated with constructing bundles using a computing tool. We have built SLIMPad, the Superimposed Layer Information Manager scratchPad, which allows problem solvers to easily select, annotate, and elaborate information from diverse information sources. SLIMPad also allows us to explore the use of superimposed information (Delcambre & Maier, 1999), maintaining active links from the SLIMPad (the superimposed layer) to the selected (referenced) information in the original sources (the base layer). SLIMPad supports freeform construction of nested bundles. Each bundle may contain one or more scraps, analogous to scraps in a scrapbook, where each scrap contains a mark or reference to the underlying selected information in its original source. At present, SLIMPad supports marks that reference a variety of source document types, including XML documents, Microsoft Excel worksheets, objects in Microsoft PowerPoint presentations, and elements in HTML pages. Examples can be viewed at http://www.cse.ogi.edu/footprints and a more complete description is available in Delcambre et al. (2000).

CONCLUSION

Field observations of expert clinicians caring for patients in intensive care units revealed widespread use of bundles as a means of managing information to support diverse, complex, and often simultaneous tasks. They appear to be especially useful for managing information in settings that are characterized by a high degree of uncertainty, a low level of predictability, and potentially grave outcomes; where time and attention are highly constrained and interruptions are frequent; and where interdisciplinary teamwork is essential. Reports of analogous observations from other fields such as aviation and air traffic control suggest that the use of bundles may be common in other domains. In an age of digital libraries, computer-based tools for creating and managing bundles may be needed as the information in these settings is increasingly represented in digital collections which promise to be much larger, more complex, more diverse, and more difficult to explore and manipulate. If successful, such tools may contribute to meeting "the traditional, and still vexing, challenge of getting critical information to those who need it in a sufficiently timely fashion that it can contribute to the quality of the decisions they make" (Scholtz, 2000).

ACKNOWLEDGMENTS

This work is supported by National Science Foundation grant NSF_9817492, part of the Digital Libraries Initiative, Phase 2 (National Science Foundation, 1998). The authors owe a special thanks to the nurses, doctors, and other clinicians who graciously shared their time, expertise, and insight.

REFERENCES

Ames, S. W. (1993). Multiple spoken and written channels of communication: An ethnography of a medical unit in a general hospital. Doctoral Dissertation, State University of New York at Buffalo. Dissertation Abstracts International, 54(02), 0369A.

Ash, J. S.; Gorman, P. N.; Hersh, W. R.; Lavelle, M.; & Poulsen, S. P. (1999). Perceptions of house officers who use physician order entry. Unpublished paper presented at the American Medical Informatics Association Fall Symposium, Washington, DC.

Delcambre, L., & Maier, D. (1999). Models for superimposed information. Unpublished paper presented at the Advances in Conceptual Modeling ER '99, Lecture Notes in Computer Science, Paris, France, November 15-18, 1999.

Delcambre, L.; Maier, D.; Bowers, S.; Deng, L.; Weaver, M.; Gorman, P.; Ash, J.; Lavelle, M.; & Lyman, J. (in press). Bundles in captivity: An application of superimposed information (Papers presented at the Seventeenth International Conference on Data Engineering, Heidelberg, Germany).

Elstein, A. S.; Shulman, L. S.; & Sprafka, S. A. (1978). Medical problem solving: An analysis of clinical reasoning. Cambridge, MA: Harvard University Press.

Evans, D. A., & Gadd, C. S. (1989). Managing coherence and context in medical problem solving discourse. In D. A. Evans & V. L. Patel (Eds.), Cognitive science in medicine: Biomedical modeling (pp. 211-256). Cambridge, MA: MIT Press.

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.

Gorman, P. N. (1995). Information needs of physicians. Journal of the American Society for Information Science, 46(10), 729-736.

Gorman, P. N. (1999). Information seeking of primary care physicians: Conceptual models and empirical studies. In T. Wilson & D. Allen (Eds.), Exploring the contexts of information seeking behaviour (Proceedings of the 2nd International Conference on Information Needs, Seeking and Use in Different Contexts, Sheffield, England, 1998) (pp. 226-240). Los Angeles: Taylor Graham.

Hutchins, E. (1995). Cognition in the wild. Cambridge, MA: MIT Press.

Hutchins, E., & Klausen, T. (1996). Distributed cognition in an airline cockpit. In Y. Engestrom & D. Middleton (Eds.), Cognition and communication at work (pp. 15-34). Cambridge, England: Cambridge University Press.

Krause, C. R.; Westdorp, J. M.; Coonen, D. A.; & Jenks, D. L. (1996). Forming an integrated documentation system. Nursing Management, 27(8), 25-26.

Krikelas, J. (1983). Information seeking behavior: Patterns and concepts. Drexel Library Quarterly, 19(2) 5-20.

Levy, D. M., & Marshall, C. C. (1995). Going digital: A look at assumptions underlying digital libraries. Communications of the ACM, 38(4), 77-84.

Mackay, W. E. (1999). Is paper safer? The role of flight strips in air traffic control. ACM Transactions on Computer-Human Interaction, 6(4), 311-340.

Marshall, C. C. (2000). The future of annotation in a digital (paper) world. In S. Harum & M. Twidale (Eds.), Successes and failures of digital libraries (Papers presented at the 35th Annual Clinic on Library Applications of Data Processing, March 22-24, 1998) (pp. 97-117). Urbana-Champaign: Graduate School of Library and Information Science, University of Illinois.

National Science Foundation. (1998). Digital Libraries Initiative. Retrieved September 27, 2000 from the World Wide Web: http://www.dli2.nsf.gov.

Newell, B. A. (1996). We killed the Kardex so the care path could live. Nursing Management, 27(12), 51.

Nygren, E., & Henriksson, P. (1992). Reading the medical record. I. Analysis of physicians' ways of reading the medical record. Computer Methods & Programs in Biomedicine, 39(1-2), 1-12.

Nygren, E.; Lind, M.; Johnson, M.; & Sandblad, B. (1992). The art of the obvious: Automatically processed components of the task of reading frequently used documents: Implications for task analysis and interface design. In P. Bauersfeld, G. Lynch, & T. Bennett (Eds.), CHI '92 conference proceedings: Striking a balance (Proceedings of the Conference on Human Factors in Computing, Monterey, CA, 1992) (pp. 235-239). New York: Association for Computing Machinery.

Osier, W. (1932). On the educational value of the medical society. In W. Osier (Ed.), Aequanimatis and other addresses to medical students, nurses, and practitioners of medicine (3d ed.) (pp. 327-345). Philadelphia: Blakiston.

Rose, D. D.; Mander, R.; Oren, T.; Ponceleon, D. B.; Salomon, G.; & Wong, Y. Y. (1993). Content awareness in a file system interface: Implementing the "Pile" metaphor for organizing information. In R. Korfhage, E. Rasmussen, & P. Willett (Eds.), SIGIR '93 Proceedings of the 16th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (Pittsburgh, PA, June 27-July 1, 1993) (pp. 260-269). New York: ACM Press.

Ruhleder, K. (1994). Rich and lean representations of information for knowledge work: The role of computing packages in the work of classical scholars. ACM Transactions on Information Systems, 12(2), 208-230.

Schmidt, H. G.; Norman, G. R.; & Boshuizen, H. P. A. (1990). A cognitive perspective on medical expertise. Academic Medicine, 65(10), 611-621.

Scholtz, J. (2000). DARPA/ITO information management program background. Retrieved September 27, 2000 from the World Wide Web: http://www.darpa.mil/ito/research/im/ background.html.

Shabot, M. M. (1997). The HP CareVue clinical information system. International Journal of Clinical Monitoring & Computing, 14(3), 177-184.

Shipman, E M.; Marshall, C. C.; & Moran, T. P. (1995). Finding and using implicit structure in human-organized spatial layouts of information. In I. R. Katz, R. Mack, L. Marks, M. B. Rosson, & J. Nielsen (Eds.), Human factors in computing systems: "Mosaic of creativity" (CHI '95: Conference Proceedings on Human Factors in Computing Systems, Denver, CO, May 7-11, 1995) (pp. 346-353). New York: Association for Computing Machinery.

Shipman, F. M., & Marshall, C. C. (1999). Formality considered harmful: Experiences, emerging themes, and directions on the use of formal representations in interactive systems. Computer Supported Cooperative Work, 8(4), 333-352.

Shipman, F. M., & McCall, R. J. (1999). Incremental formalization with the hyper-object substrate. ACM Transactions on Information Systems, 17(2), 199-227.

Simon, H. A. (1955). A behavioral model of rational choice. Quarterly Journal of Economics, 69(1), 99-118.

von Thaden, T. (2000). Social informatics and aviation technology. Bulletin of the American Society for Information Science, 26(3), 13-14. Digital Libraries and Their Challenges

PAUL GORMAN, M.D., is currently an Assistant Professor in the Division of Medical Informatics and Outcomes Research at the Oregon Health Sciences University and Assistant Director of Medical Education at Providence Portland Medical Center in Portland, Oregon. Dr. Gorman completed his formal training at Rush Medical College in Chicago and learned about the realities of primary care while practicing in an Oregon coastal community. His research, funded by the National Science Foundation and the National Library of Medicine, is focused on information seeking and information management by expert clinicians.

JOAN ASH teaches in the Division of Medical Informatics and Outcomes Research in the School of Medicine, Oregon Health Sciences University. Formerly Associate Director of the library there, her interest in digital libraries research has user information needs as a focus. In related work, she heads a team exploring computerized physician order entry using qualitative methods.

MARY LAVELLE is a Research Associate in the Division of Medical Informatics and Outcomes at the Oregon Health Sciences University. She worked as a staff nurse for a number of years and developed an interest in the history of care and care rituals. Other research interests include communication/information sharing in care delivery and the use of qualitative methods in health services research. She holds an M.S. in nursing from Oregon Health Sciences University.

JASON LYMAN, M.D., is an Assistant Professor in Clinical Informatics at the University of Virginia. His research interests include physician order entry, consumer informatics, clinical data warehousing, and multimedia information management in medicine. He is an avid photographer with an interest in QuickTime Virtual Reality applications for medicine. Prior to his current position, Dr. Lyman completed a National Library of Medicine Postdoctoral Fellowship in Informatics at Oregon Health Sciences University, where he also earned a master's degree in medical informatics. He is a member of both the American Medical Informatics Association and the American Academy of Pediatrics.

LOIS DELCAMBRE is currently a Professor of Computer Science and Engineering at the Oregon Graduate Institute in Beaverton, Oregon. She works in the database field of computer science with a particular interest in data models and conceptual modeling. She and David Maier are developing generic technology to manage superimposed information to support a wide range of superimposed applications. She has published over 70 papers and been a principal investigator in a wide range of research projects. Dr. Delcambre received her Ph.D. in Computer Science from the University of Louisiana, Lafayette, Louisiana, in 1982, her M.S. in Mathematical Sciences from Clemson University in 1974, and her B.S. in Mathematics from the University of Louisiana, Lafayette, Louisiana, in 1972.

DAVID MAIER is a professor of Computer Science and Engineering at the Oregon Graduate Institute of Science and Engineering. He holds a Ph.D. degree in Electrical Engineering and Computer Science from Princeton University and a B.A. degree with a double major in Mathematics and Computer Science from the University of Oregon. His research interests include object-oriented database systems, query processing, database support for scientific computing, semi-structured data, and superimposed information systems.

MATHEW WEAVER is a second year doctoral student in the Department of Computer Science and Engineering at the Oregon Graduate Institute of Science and Technology. He received a Bachelor of Science degree in Computer Science from Brigham Young University in 1999. His research interests include architectures for superimposed information systems and information systems for forest management. He also has experience with information management for environmental monitoring through his work as a consultant to EarthSoft, Inc.

SHAWN BOWERS is a third year doctoral student in the Department of Computer Science at the Oregon Graduate Institute of Science and Technology. He has a bachelors of science degree in Computer and Information Science from the University of Oregon. His current research interests center around exploring the structural aspects of data models by applying database technology to various loosely structured models like Topic Maps, RDF, and XML; developing technology to support superimposed information; and defining techniques for information integration and transformation. He is a member of the Tracking Footprints Through an Information Space, Harvesting Information to Sustain Our Forests, and Autonomix research projects.

Paul Gorman, Joan Ash, Mary Lavelle, Jason Lyman, Oregon Health Sciences University, 3181 Sam Jackson Park Road, Portland, OR 97201-3098

Lois Delcambre, David Maier, Mathew Weaver, Shawn Bowers, Oregon Graduate Institute of Science and Technology, 20000 N.W. Walker Road, Beaverton, OR 97006