42 1094-7167/02/$17.00 © 2002 IEEE IEEE INTELLIGENT SYSTEMS

H C C a t N A S A

Designing Human-Centered DistributedInformation SystemsJiajie Zhang, University of Texas at Houston

Vimla L. Patel, Columbia University

Kathy A. Johnson and Jack W. Smith, University of Texas at Houston

and NASA Johnson Space Center

Jane Malin, NASA Johnson Space Center

Many computer systems are designed according to engineering and technology

principles and are typically difficult to learn and use. The fields of human-

computer interaction, interface design, and human factors have made significant contri-

butions to ease of use and are primarily concerned with the interfaces between systems

and users, not with the structures that are often morefundamental for designing truly human-centered sys-tems. The emerging paradigm of human-centeredcomputing (HCC)—which has taken many forms—offers a new look at system design.

HCC requires more than merely designing an arti-ficial agent to supplement a human agent. Thedynamic interactions in a distributed system com-posed of human and artificial agents—and the con-text in which the system is situated—are indispens-able factors. While we have successfully applied ourmethodology in designing a prototype of a human-centered intelligent flight-surgeon console at NASAJohnson Space Center, this article presents a method-ology for designing human-centered computing systems using electronic medical records (EMR)systems.

Distributed cognitionWe base our human-centered computing per-

spective on the theory of distributed cognition.1–9

According to this theory, people behave in infor-mation-rich environments filled with natural objects,artificial objects, and agents, whether human or non-human. This environment, according to the theory,is grounded in elaborate social and cultural struc-tures. In everyday life, people must process infor-mation derived from human agents, artificial agents,and groups of agents. It is the interwoven process-ing of such information that generates intelligentbehavior.

Distributed cognition considers human and artifi-cial agents to be indispensable components of a sin-gle distributed system.5 Human activities in concretesituations are guided, constrained, and even deter-mined by the physical and social context in whichthey operate.9,10 One study has argued that the prop-erties of a distributed cognitive system consisting ofa group of human agents interacting with complexsystems—in an airplane cockpit or the control roomof a ship—can differ radically from the properties ofthe components, and they cannot be inferred fromthe properties of the components alone, no matterhow detailed the knowledge of the properties of thosecomponents might be.1,11

The theory of distributed cognition makes threemajor claims:

• The unit of analysis is the system composed ofhuman and artificial agents.

• The pattern of information distribution amonghuman and artificial agents can radically changethe behavior of the distributed system.

• The behavior of a distributed system should bedescribed by the information flow dynamics.

One research project described a similar system asthe triples rule, which states that the unit of analysisfor cognitive engineering and computer science is atriple composed of person, machine, and context.12

The goal of a traditional approach is to make userinterfaces transparent so users can completely engage

Human-centered

distributed information

design methodology

incorporates the theory

of distributed cognition

with the need for

multiple levels of

analyses in system

design. This

methodology provides

systematic principles,

guidelines, and

procedures for designing

human-centered

computing systems.

in their primary tasks. With transparent inter-faces, users directly interact with the system.However, user-friendly interfaces are far fromsufficient as a design goal. In our methodol-ogy, user-friendly interfaces are only one offour levels of analysis. The three other lev-els—functions, tasks, and users—are just asimportant. For example, an excellent user-friendly interface might be irrelevant if thetask the interface supports is not the task theuser performs.

Design methodologyOur methodology, called human-centered

distributed information design (HCDID),incorporates the theory of distributed cogni-tion with the need for multiple levels ofanalyses in system design. We developed thismethodology to provide systematic princi-ples, guidelines, and procedures for design-

ing HCC systems. We selected electronicmedical records systems to demonstrate ourmethodology because HCC is almost nonex-istent in EMR systems.

EMR systems are highly complex, dis-tributed information systems that, withproper human-centered design, have thepotential to improve the quality of health caredramatically. The lack of even minimum con-siderations of HCC design principles in mostcurrent EMR systems makes them very dif-ficult to learn and use. In turn, this difficultyleads to strong resistance by physicians, andin some cases it leads to abandoning EMRsystems altogether. The EMR system we dis-cuss here is not a specific system. Rather, itis a collection of EMR systems that are cur-rently on the market.

Figure 1 shows the HCDID methodology,which consists of three major components.

The components on the left are multiple lev-els of analyses for single-user, human-cen-tered design. The functional, task, and repre-sentational analysis levels have degrees ofabstraction—with the level of functionalanalysis most abstract and the level of repre-sentational analysis most concrete. The user-analysis level contributes to each of the lev-els of functional, task, and representationalanalysis. The components on the right repre-sent the additional analysis needed for design-ing distributed human-centered systems.

User analysisUser analysis provides user information to

the functional, task, and representationalanalyses. User analysis is the process of iden-tifying user characteristics, such as expertise,skill, knowledge, educational background,cognitive capacity, and so forth. User analy-

SEPTEMBER/OCTOBER 2002 computer.org/intelligent 43

User analysis: Identify characteristics of userssuch as expertise and skills, knowledge base, age,

education, cognitive capacities and limitations,perceptual variations, and so on.

Functional analysis: Identify top-leveldomain structure and ideal task space

independent of implementations

Distributed user analysis: Analyze division oflabor, overlap of knowledge and skills,

pattern of communication, and social interaction

Multiple levels of analysis Distributed cognition analysis

Distributed functional analysis: Analyze top-levelinterrelations and constraints of human and artificial

agents in the domain

Distributed task analysis: Analyze space and timedistributions of activities, information flow dynamics

across human and artificial agents, and compatibilitiesbetween agents

Distributed representational analysis: Analyze how thesame information can be distributed across human and

artificial agents in different ways under differentimplementations, and find out which implementation is

most efficient for which task

Task analysis: Identify system functions that must be performed, procedures and actions to be

carried out, information to be processed, patternand dynamics of information flow, input and

output formats, constraints that must beconsidered, communication needs, and the

organization and structure of the task

Representational analysis: Identify thebest information display and the best

information flow structure for a given tasksuch that the interaction between users and

systems is in a direct interaction mode

Contents for system implementation:Functional requirements; task structures; information flow

diagrams; task-specific, event-related, and context-sensitiveinformation displays; interface design

recommendations; and so on

Figure 1. The methodology of human-centered distributed information design. Each of the four distributed analyses contributes to its corresponding analysis on the left. The component at the bottom represents the product of functional, task, and representational analyses. For each level of analysis, our general methodology lets us employ several alternative, specific methods.

sis is necessary for designing systems thathave an information structure that mustmatch user knowledge levels.

For an EMR system, different users oper-ate different system components. These usersinclude installers, maintainers, administra-tors, nurses, physicians, registration person-nel, laboratory technicians, billing staff, andpatients. Different users will have differentlevels of understanding when it comes to thesame component of the system.

Researchers have written about the impor-tance of technology in shaping cognitiveactivity and the impact of cognition on tech-nology design.

13These studies show that

using EMR systems requires several changesin physicians’and trainees’ information-gath-ering and reasoning strategies. For example,there are differences in the content and orga-nization of information, with paper recordsorganized according to a narrative structureand computer-based records organizedaccording to discrete items of information.This difference in knowledge organizationresults in different data-gathering strategies,where the doctor-patient dialogue is cogni-tively influenced by the structure of the EMRsystem.

For distributed human–computer systems,it is critical to analyze users’ group proper-ties, such as division of cognition and activ-ity, overlap of knowledge and skills, com-munication channels, social status, and so on.For example, whether two human agents arebetter than one human agent often dependson the overlap of knowledge and skills. Theways in which effort is distributed across agroup of human agents can significantlyaffect the outcome of a task.

A recent study shows the importance ofproper user analysis for EMR systems.14

Nurses and physicians—who have partiallyoverlapping knowledge bases, skills, expe-riences, and job responsibilities—rely ondifferent patient strategies that can lead themto different understandings, diagnoses, andsubsequent activities. Another study char-acterized the qualitative nature of team inter-action and its relationship to training healthprofessionals in a primary care unit in aBoston hospital.15 This study showed thatthe demarcation of responsibilities and rolesof personnel within the team can becomefuzzy.

The nature of the individual expertiserequired was a function of the patient prob-lem and the interaction goal. Team charac-teristics often contribute to reduction of

unnecessary and redundant interactions.Distributed responsibilities let the teamprocess massive amounts of patient infor-mation, which reduces individual cognitiveloads. Individual expertise contributes toaccomplishing team goals, although teamperformance shows a completely differentstructure.

You can obtain user information with indi-rect or direct methods. With direct methods,you would collect the information from theusers with surveys, questionnaires, field vis-its, focus groups, interviews, or any of severalethnographic observation techniques. Withindirect methods, you collect the informationfrom other sources, such as textbooks, hand-books, protocols, procedures, job descrip-

tions, operation rules, and other training andeducation materials.

Functional analysisFunctional analysis is the process of iden-

tifying critical top-level domain structures—goals that can be largely independent ofimplementations. Functional analysis is moreabstract than task and representational analy-ses because it does not involve details of taskprocesses and representation. In other words,functional analysis is similar to traditionalrequirements analysis. Work domain analy-sis (which focuses on the structures of a workdomain) and cognitive work analysis (whichfocuses on cognitive activities in the workdomain) are two of the methodologies thatcan help in conducting functional analysis.For a distributed system, functional analysisalso identifies a system’s artificial and humanagents, their interrelations and constraints,and their essential roles. For a knowledge-rich domain such as medicine or aviation,functional analysis requires the acquisition

of detailed domain knowledge and a deepunderstanding of domain structures.

Consider one of the products of doing afunctional analysis for an EMR system. Anideal EMR should be able to support the fol-lowing functions: data, alerts, reminders,schedules, clinical decision support, medicalknowledge, communication, and other aids.These functions should be complete, accu-rate, and timely. These functions should alsobe available for all types of health-care pro-fessionals. Furthermore, these functionsshould be available at all times and at allpoints of care. The ideal EMR should over-come the known problems associated withpaper-based records but still provide thatkind of functionality. Finally, and mostimportantly, an ideal EMR should be able toimprove the quality of health care.

A distributed functional analysis of EMRcan show the variety of artificial and humanagents and their interrelations. Examples ofartificial agents include various moduleswithin EMR, such as patient record module,alert module, decision-support module, andcommunication module. Examples of humanagents include physicians, nurses, laboratorytechnicians, and registration staff. An exam-ple of the interrelations between artificial andhuman agents is a mapping that shows whichartificial agent supports which human agent,and to what degree.

Task analysisTask analysis is more concrete than func-

tional analysis because it involves specific taskstructures and procedures. However, it is stillmore abstract than representational analysisbecause it does not involve the details of howinformation is represented. Task analysis is acritical component in cognitive systems engi-neering and usability engineering. It consists ofthe process of identifying system functions,task procedures, input and output formats, con-straints, communication needs, organizationstructures, information categories, and taskinformation flow.

One important function of task analysis isensuring that the system implementationincludes only the necessary and sufficient taskfeatures that match user capacity and arerequired by the task. Fancy features—and fea-tures that do not match user capacity or are notrequired by the task—might only generateadditional processing demands for the user,and thus make the system harder to use. Butavoiding unnecessary features does not meanexcluding adaptation mechanisms that dynam-

44 computer.org/intelligent IEEE INTELLIGENT SYSTEMS

H C C a t N A S A

One important function of task

analysis is ensuring that the system

implementation includes only the

necessary and sufficient task

features that match user capacity

and are required by the task.

ically adjust the interactions between usersand tasks in changing contexts.

Hierarchical task analysis, which is thebasic method for revealing the structure ofany task that has goals and subgoals,describes a task in terms of a hierarchy ofoperations by using a graphical representa-tion. Hierarchical task analysis decomposesa high-level task into its constituent subtasksand operations, and is essential for any sys-tem consisting of both human and artificialagents. This type of analysis typically showsthe big picture of the task components andthe relations among them. It also takes intoaccount both physical and mental activities.

The theory of distributed representations isthe foundation for the analysis of the distri-bution of information patterns among humanand artificial agents.5,6 Human agents inter-act with each other and with artificial agentssynchronously and asynchronously—at oneplace and at different places. The analysis oflocations and activities can reveal how tasksare distributed across space and over time.

One study demonstrated this type of analy-sis on more than 20 medical personnel, whointeracted over the Internet and through con-ference calls over a period of two years. Thepurpose of the study was to develop softwarethat could be used in any hospital system.16

Information flow analysis can help analyzehow the information gets propagated andtransformed among human and artificialagents. Distributed task analysis, by the sametoken, can reveal critical task structures thatcannot be identified by conventional taskanalysis, which focuses on a single individ-ual’s interaction with a system.

Task analysis can help identify task struc-tures, interaction among procedures, andinformation flow. For example, task analysiscan identify overlooked tasks, relative taskimportance, overlapping task information,function groups, relation to user analysis, andso on. It can also help pinpoint task bottle-necks where special design requirementscome into play. Another product of taskanalysis is a taxonomy based on information-processing needs. For example, there areinformation tasks for retrieval, gathering,seeking, encoding, transformation, calcula-tion, manipulation, comparison, organization,navigation, and so on. Identifying differentinformation-processing needs is essential forthe creation of task-specific, context-sensi-tive, and event-related information displays.

A systematic and comprehensive taskanalysis of an EMR system is out of reach

for any individual or even a small researchgroup. However, it is essential for designinga new EMR system or redesigning an exist-ing EMR system to make it human-centered.Commitment from a large institution isessential for a comprehensive task analysis.We have carried out preliminary task analy-sis of several small components of EMR sys-tems. For example, one analysis showed thatwriting a prescription using an EMR systemrequires more steps and much more time thanwriting a paper prescription, although themental effort required is less for EMR thanfor paper prescription writing.

Additionally, performing a hierarchical taskanalysis to identify an EMR system’s under-lying data structure can help pinpoint some

fundamental problems. For example, one studyshowed that current EMR systems use one oftwo data structures, neither of which is drivenby human-centering.13 One EMR data struc-ture, for example, uses a hierarchical datamodel to capture information used by specificapplications. It is primarily a patient-recordsystem added onto a billing system. The otherdata structure makes extensive use of an event-based approach by recording informationaccording to a time-oriented view to facilitateits reuse by multiple applications. Unfortu-nately, these two data structures do not supportthe daily tasks of health-care professionals, sothey are not human-centered. A typical dailytask, such as making a diagnosis, is better sup-ported by an EMR system that organizes infor-mation around problems.

Because EMR systems are used by manytypes of people—synchronously and asyn-chronously, and at the same or different loca-tions—a distributed task analysis can showcrucial interactions among human and artifi-cial agents and can provide an understanding

of these interactions for designing accordingto human-centered principles. To our knowl-edge, very little analysis of this type has beendone for EMR systems.

Representational analysisFunctional analysis and task analysis deal

with system functions, structures, and pro-cesses, whereas representational analysisdeals with the interface between systems andusers. Representational analysis is based ona robust phenomenon called representationaleffect,5 in which different representations ofa common abstract structure or process cangenerate dramatically different representa-tional efficiencies, task difficulties, andbehavioral outcomes.

The form of a representation can influenceand sometimes determine what informationcan be easily perceived, what processes canbe activated, and what information can bederived from the representation. For a com-plex or novel task, some portion of the taskspace might never be explored, and somestructures of the task might never be discov-ered without a change in representation. Rep-resentational analysis can be performed onsystem properties that are identified throughfunctional and task analysis. With direct-interaction interfaces, users can efficientlyengage in the primary tasks they intend toperform. And they can avoid the interfacehousekeeping tasks that typically act as bar-riers between users and systems.

Representational analysis can help sys-tematically generate innovative ideas for datadisplays and information for exploration,evaluation, and selection. This type of analy-sis works through developing a taxonomy of displays and tasks, and through map-ping principles between displays and tasks.Researchers have applied representationalanalysis to relational information displaysand even cockpit instrumentation. A repre-sentational taxonomy of relational informa-tion displays, when combined with a tasktaxonomy from task analysis, can systemat-ically determine the best display format fora specific task.

Representational analysis can help identifyalternative displays for each task in the taxon-omy of information-processing needs and canalso help determine the best match between adisplay and a task for each unique event undereach unique situation. This type of analysis thendetermines the mapping between a display anda task with a mapping principle: the displayshould carry exact information for the task, no

SEPTEMBER/OCTOBER 2002 computer.org/intelligent 45

Distributed task analysis can show

crucial interactions among human

and artificial agents and can

provide an understanding of these

interactions for designing according

to human-centered principles.

more and no less. The task-specific, context-sensitive, and event-related displays are basicelements for implementing HCC systems.

Medical records can be organized anddisplayed in different ways in EMR sys-tems. Source-oriented displays present dataorganized by the data source, such as labreports, radiology reports, or physical exam-inations. Concept-oriented displays groupdata according to clinical problems. Forexample, a concept-oriented display forabdominal pain might present history, radi-ology reports, blood tests, and assessmentsin a single display.

Medical data can also be organized and dis-played along timelines and according to dif-ferent processes. Several studies have foundthat different types of displays are good fordifferent tasks and that no single type of dis-play is good for all tasks for all users. Thisfinding is the mapping principle between tasksand representations. Empirical findings arecertainly important, but it is neither practicalnor necessary to conduct empirical testing forevery system component.

The representational analysis, which isbased on established theories, principles, andgeneralizable empirical regularities, can offer

more than the empirical studies. When com-bined with task, user, and functional analy-ses, representational analysis can systemat-ically generate human-centered displays andtasks for several users, which is essential fordesigning human-centered EMR systems.

The products generated from ourHCDID methodology are the contents

for implementing human-centered distributedcomputing systems. Examples of these prod-ucts are functional requirements, descriptionsof task structures and procedures, specifica-tions of information-flow dynamics, andideas for task-specific, event-related, and con-text-sensitive information displays. Althoughother cognitive engineering methods mightgenerate the same contents, HCDID offersunique perspective, principles, and proce-dures that are rooted in the theory of distrib-uted cognition and an approach that requiresmultiple levels of analysis.

HCDID considers HCC not just at the lev-els of representations, but also at the levels ofusers, functions, and tasks. In many systemdesigns thought to be human-centered,human-centered principles are mainly appliedat the representation level. A poor display ofa good task structure that has the same func-tions of the same domain might be muchmore efficient than an excellent display on apoor task structure of the same domain. Dis-plays are also critically dependent on users. Agood display for one user might be a poor dis-play for a different user because of user vari-ation. As a general principle in human-cen-tered display design, there will be no idealdisplay for all users in every context.

We are in the process of evaluating ourmethodology. We plan to refine it to makedesigning any new distributed, human-com-puter system much easier.

AcknowledgmentsThis article was supported in part by grant No.

NCC 2-1234 from NASA’s Human-CenteredComputing and Intelligent Systems Program.

46 computer.org/intelligent IEEE INTELLIGENT SYSTEMS

H C C a t N A S A

T h e A u t h o r sJiajie Zhang is an associate professor at the University of Texas at Houston.His research interests include human-computer interaction, cognitive sci-ence and engineering, distributed cognition, and health informatics. Hereceived a PhD in cognitive science from the University of California at SanDiego. Contact him at the School of Health Information Sciences, Univ. ofTexas, Houston, 7000 Fannin, Ste. 600, Houston, TX 77030; jiajie.zhang@uth.tmc.edu.

Vimla L. Patel is a professor of medical informatics and medical psychol-ogy at Columbia University. Her research interests include medical cognition,cognitive evaluation of human-computer interaction, collaborative health-care teams, and naturalistic decision making in critical-care environments. Shereceived a PhD in educational psychology from McGill University in Mon-treal, Canada. Contact her at the Dept. of Medical Informatics, ColumbiaPresbyterian Medical Center, Vanderbilt Clinic 5, 622 West 168th St., NewYork, NY 10032-3720; patel@dmi.columbia.edu.

Kathy A. Johnson is an assistant professor at the University of Texas atHouston and the medical informatics lead at NASA Johnson Space Center.Her research interests include decision-support systems, intelligent tutoringsystems, human-centered computing, and databases. She received a PhD incomputer science from Ohio State University. Contact her at the School ofHealth Information Sciences, Univ. of Texas, Houston, 7000 Fannin, Ste. 600,Houston, TX 77030; kathy.a.johnson@uth.tmc.edu.

Jane Malin is the technical assistant to the branch chief in the IntelligentSystems Branch at NASA Johnson Space Center. Her research interestsinclude human-centered computing, methods for engineering advanced sup-port software, autonomous monitoring, control agents, and intelligent mod-eling and simulation. She received a PhD in experimental psychology fromthe University of Michigan. Contact her at Mail Code ER2, NASA JohnsonSpace Center, Houston, TX 77058-3696; jane.t.malin1@jsc.nasa.gov.

Jack W. Smith is a professor at the University of Texas at Houston, and thedeputy director of medical informatics and health-care systems at NASAJohnson Space Center. His research interests include modeling problem-solving in health care, the representations of knowledge for automating theseprocesses, and the application of cognitive science to the understanding ofhuman–computer interaction. He received an MD in clinical pathology fromWest Virginia University Medical School and a PhD in computer sciencefrom Ohio State University. Contact him at the School of Health Informa-tion Sciences, Univ. of Texas, Houston, 7000 Fannin, Ste. 600, Houston, TX77030; jack.w.smith@uth.tmc.edu.

ReferencesReferences1. E. Hutchins, Cognition in the Wild, MIT

Press, Cambridge, Mass., 1995.

2. D.A. Norman, Things that Make Us Smart,Addision-Wesley, Boston, Mass., 1993.

3. G. Solomon, Distributed Cognitions: Psycho-logical and Educational Considerations, Cam-bridge Univ. Press, Cambridge, U.K., 1996.

4. P.C. Wright, R.E. Fields, and M.D. Harrison,“Analyzing Human-Computer Interaction asDistributed Cognition: The ResourcesModel,” Human-Computer Interaction, vol.15, no. 1, 2000, pp. 1–41.

5. J. Zhang and D. A. Norman, “Representationsin Distributed Cognitive Tasks,” CognitiveScience, vol. 18, no. 1, Jan./Mar. 1994, pp.87–122.

6. J. Zhang, “A Distributed RepresentationApproach to Group Problem Solving,” J. Am.Soc. of Information Science, vol. 49, no. 9,July 1998, pp. 801–809.

7. J. Holland, E. Hutchins, and D. Kirsh, “Dis-tributed Cognition: Toward a New Founda-tion for Human-Computer InteractionResearch,” ACM Trans. Computer-HumanInteraction, vol. 7, no. 2, June 2000, pp.174–196.

8. D.D. Woods, “Designs Are Hypotheses AboutHow Artifacts Shape Cognition and Collab-oration,” Ergonomics, vol. 41, no. 2, Feb.1998, pp. 168–173.

9. W. J. Clancey, Situated Cognition, LawrenceErlbaum, Mahwah, N.J., 1997.

10. R.R. Hoffman et al., “The Triples Rule,” IEEEIntelligent Systems, vol. 17, no. 3, May–June2002, pp. 62–65.

11. L. Suchman, Plans and Situated Actions,Cambridge Univ. Press, Cambridge, U.K.,1987.

12. E. Hutchins, “How a Cockpit Remembers ItsSpeed,” Cognitive Science, vol. 19, no. 3,July–Sept. 1995, pp. 265–288.

13. V.L. Patel et al., “Impact of a ComputerizedPatient Record System on Medical Data Col-

lection, Organization, and Reasoning,” J. Am.Medical Informatics Assoc., vol. 7, no. 3,July–Sept. 2000, pp. 569–585.

14. C.M. Johnson et al., “Understanding the Men-tal Models of Nurses and Physicians ReadingPatient Records,” Proc. AMIA, Am. MedicalInformatics Assoc., Washington, D.C., 2002.

15. V.L. Patel et al., “The Collaborative HealthCare Team,” Teaching and Learning in Med-icine, vol. 12, no. 3, Summer 2000, pp.117–132.

16. V.L. Patel et al., “Toward a Framework forComputer-Mediated Collaborative Design inMedical Informatics,” Methods of Informa-tion in Medicine, vol. 38, no. 3, Sept. 1999,pp. 158–176.

17. J.J. Cimino et al., “What is Wrong withEMR?,” Proc. Am. Medical InformaticsAssoc., Washington, D.C., 1999.

For more information on this or any other com-puting topic, please visit our Digital Library athttp://computer.org/publications/dlib.

SEPTEMBER/OCTOBER 2002 computer.org/intelligent 47

Let your e-mail address show your professional commitment.An IEEE Computer Society e-mail alias

forwards e-mail to you, even if you change

companies or ISPs.

you@computer.org

The email address of computing professionals