The visible and the invisible: Distributed Cognition for medical devices Dominic Furniss, Ann Blandford, Atish Astrid Mayer Rajkomar & Chris Vincent Department of Oncology UCLIC Royal Free NHS Trust UCL, Gower Street Pond Street London WC1E 6BT, UK London, NW3 2QG, UK +44 20 7679 0688 a.blandford@ucl.ac.uk provoke further discussion – e.g. the cases of Denise ABSTRACT Melanson [10] and Lisa Norris [16]. However, such high Many interactive medical devices are less easy to use than profile incidents are mercifully rare, and many incidents are they might be, and do not fit as well as they could in their minor and may not be reported at all. For example, Husch contexts of use. Occasionally, the deficiencies lead to et al. [7] suggest that few incidents are reported. In a study serious incidents; more often, they have a less visible effect of infusion pump use in a busy hospital, covering 426 on the resilience and efficiency of healthcare systems. intravenous infusions, they identified a total of 389 errors, These issues remain largely invisible as they are not occurring in 285 of the infusions. In other words, 2/3 of the reported and have rarely been studied. In this paper, we infusions on which data was gathered involved at least one report on the use of DiCoT as an approach to representing error. Many of these errors would be classed as minor, but and reasoning about medical work, and about the role of 55 were either rate deviation or incorrect medication errors, device design within that work. We focus in particular on which had the potential to be serious. For comparison, only the design and use of infusion devices. This work highlights 48 incidents in the same categories had been reported the value of observational studies for engineering through the formal reporting system over the previous two interactive medical devices, and illustrates the value of a years from the same hospital. As discussed below, it might systematic approach to gathering and analyzing qualitative have been inappropriate to class all 389 events as “errors”, data. but this study highlights what a small proportion of errors are reported. Keywords Distributed Cognition, medical devices, DiCoT, situated However, error cases alone are not sufficient to engineer interaction, infusion devices. good systems: it is also necessary to have a good understanding of normal practice. In this paper, we present INTRODUCTION a study of normal practice in an oncology day care unit, To improve the engineering of interactive medical devices, focusing particularly on the use of infusion devices. We use it is essential to understand how those devices are used in DiCoT (Distributed Cognition for Teamwork) [3] as a context, as well as considering the engineering of the framework for structuring observations and to support devices in isolation (e.g. ensuring consistency, reliability reasoning about design. We illustrate modes of reasoning and safety of interactions). In this paper, we focus on the about design by discussing two design requirements that use of infusion devices, relatively simple devices that are were identified in our studies. used by both clinical professionals and lay people, but particularly by nurses. The use of such devices is inherently DISTRIBUTED COGNITION complex: even if the devices are configured as simply as Distributed Cognition has emerged as an approach to possible, they are used in a variety of environments, as part reasoning about system design that starts from the premise of a complex set of tools and procedures. that the ways that people make decisions and interact are dependent on the external environment as well as internal One source of information about the impact of device cognitive processes: that the environment provides design on use is to be found in incident reports, particularly resources to support thinking [5]. Furthermore, the structure root cause analyses, such as the reports in the MAUDE of the environment can be analyzed from a cognitive (Manufacturer and User Facility Device Experience) perspective; i.e. the people, roles, tasks, artifacts and the database [12]. Occasionally, incidents hit the headlines and physical layout of the system will impact the way information is processed. For example, a bridge of a ship Copyright © 2011 for the individual papers by the papers' [8] and an aircraft cockpit [9] have been analysed in this authors. Copying permitted only for private and academic way. Distributed Cognition therefore describes how socio- purposes. This volume is published and copyrighted by technical systems are structured to process information. the editors of EICS4Med 2011. 1 Properties of the system that help or hinder the processing connects the patient to the bag. These devices can be of information can then be identified and engineered. programmed at specified volumes, times and rates. The Distributed Cognition has been applied as an approach to interface on the pump broadly consists of a number entry understanding healthcare systems; for example, Nemeth et system and a display. al [13] and Xiao [18] analyse the roles of artifacts in Infusion pumps are commonly configured for the different supporting communication within clinical teams. However, needs of intensive treatment units, paediatrics units and the focus of these studies has been on facilitating more general wards. This study focuses on an Oncology communication rather than supporting the situated work of Day Care Unit. The unit provides treatment to patients on a an individual nurse, or reasoning about the design of a day basis, i.e. typically patients will come in, get treatment particular device. and return home on the same day. This includes the use of Distributed Cognition (DC) has traditionally involved a infusion pumps for intravenous treatment; e.g. high degree of craft skill on the part of the analysts. Two chemotherapy treatment. different approaches to codifying DC have been proposed. Due to their wide use and importance it should be no Wright et al [17] present the Resources Model as a surprise that others have studied the broader class of structured approach to reasoning about the design of an infusion pumps. Lin et al. [11] assessed a PCA (patient- interactive computer system from a DC perspective, controlled analgesia) pump, identified HCI issues and focusing on the resources that the system makes available to proposed a redesign with a lower likelihood for error. its user. The Resources Model approach is tailored to the Obradovich and Woods [15] evaluated a syringe pump that analysis of individual human–computer interactions. In patients take home to use. Through interviews and contrast, the Distributed Cognition for Teamwork (DiCoT) evaluation, they found complex sequences, mode [3] approach focuses attention on interactions between confusions and arbitrary alarms that needed redesigning. multiple people and multiple artifacts, and how the design More recently, pro-formas have been proposed to of technology influences those interactions. A DiCoT standardise the observation of infusion pump use [1]; and analysis involves constructing five interdependent models: nurses’ acceptance of infusion pump use with error- information flow, physical, artefact, social and reducing software has been studied [2]. Our study took an evolutionary. These models each have associated principles exploratory approach to investigate HCI issues with from the distributed cognition literature. The method volumetric infusion pumps in use in the Day Care Unit provides a structured approach for engaging with socio- (DCU). To our knowledge the two issues we highlight have technical systems. In the study reported here, we focus on not been reported elsewhere. the use of DiCoT to reason about the design of infusion METHOD pumps. Data for this study were gathered by conducting Furniss and Blandford [4] identify four ways in which observations in the DCU. In addition, two members of staff DiCoT can assist in moving from analysis to design and in the unit were interviewed to clarify issues that had arisen engineering: in the observations. For the observations, extensive field 1. To explain the basic mechanics of a system, e.g. so notes were taken, structured according to the themes of DC. its structure and functions are understood. Interviews were audio recorded and transcribed. Data gathering lasted for 5 days. These were spread over a 2. The development of deep conceptual insight, e.g. number of weeks to allow for reflection between data we found the property of ‘buffering’ is particularly gathering days. Our primary focus was on the design and important to the performance of ambulance use of infusion pumps. A secondary focus was to dispatch [3]. understand the context in which they are used. Here we 3. Identifying opportunities for incremental focus on how the pumps were set up and used. developments to improve the system. We focused on the information flow, physical and social 4. Considering revolutionary designs where the models – to build an understanding of the infusion pump system may work in a fundamentally different way. programming task and the environment in which they In this paper we focus on two incremental design worked. We gathered data to describe the system in terms considerations from disturbances that were observed in of the models, and used the associated principles to help practice. embellish this picture. Disturbances in performance were noted in conjunction with direct observations and by BACKGROUND: INFUSION PUMPS interrogating the developing models. The models’ Infusion pumps are important ubiquitous devices in representations would often crystallise observations and hospitals. Volumetric infusion pumps are typically used to raise questions that would need further data gathering. pump nutrients or medications from bags into patients intravenously. They control the rate of fluid in the line that 2 Figure 1: Task steps and disturbances in infusion pump interaction OBSERVATION RESULTS: NORMAL WORK x The pump displays zero values for the VTBI. The 31 programmable infusion interactions were observed over value needs to be entered by the user before pressing OK the 5 days; not all observations were complete because key to confirm the value. presses were not always visible. The nurses’ interactions x The nurse can then enter either the time or infusion were often very fast and without error or issue. rate. Once they have confirmed either of these values by We first describe the normal stages of setting up a pump, pressing OK, the pump calculates the missing value; i.e., and then describe two of the disturbances that were if the pump knows the VTBI and time it can work out the observed. The normal stages for programming an infusion rate, and if the pump knows the VTBI and the rate it can pump, which we observed in most cases, are as follows (see work out the time. Figure 1): x Once all these values have been checked, the user x The pump is turned on. presses the START button and the infusion commences. x The eject button is pressed to open the pump’s door. Over the course of the observation period, several kinds of The tube that connects the bag to the patient is inserted disturbance to this normal flow of activities were observed. and the door is closed. Here, we discuss two of them. x The pump asks the user to release the roller clamp and VTBI (Volume To Be Infused) issue press OK when they have done so. The roller clamp’s This issue relates to the stage in programming the infusion release allows the fluid to flow from the bag to the pump that needs the VTBI value. It is the first value that is patient. required by the pump; it is a stage that cannot be skipped, and sometimes nurses do not have this value so it needs to be calculated manually. This is noted as disturbance 5 in Figure 1. 3 As well as specifying the type of medication, the DISCUSSION prescription should detail the VTBI, the infusion rate and We have presented an example of normal work and two the time. However, in the incident that drew our attention to disturbances to that work (drawn from a larger set, to this issue, this was not the case. In this incident, the illustrate the roles of observation and structured analysis in observer (hereafter referred to as A1) observed a nurse informing design). The description of normal work, which interact with the pump far more than normal. A1 overheard forms a basis for part of the DC analysis of nurses’ work in the nurse tell the patient that maths was not their strong the DCU, could, in principle, have been based on point to make conversation and to allude to the difficulty documentation of how to use the device, but was validated they were having. A1 observed the nurse turn the pump on through observations of nurses at work. The disturbances and off, and then program the pump with little difficulty. that we observed are undocumented, and can only be The nurse was too busy to discuss the matter at the time but identified through observation. They are not sufficiently we later found out the VTBI was not on the prescription disruptive to feature in incident reports, and therefore chart and so they had to work it out mentally. would not be identified if incident reports were the major The prescription instructed the nurse to set up an infusion source of information to inform new design; nevertheless, with a rate of 15ml/hr over a 20 minute period. This is a they are significant disturbances to normal work, and standard calculation a nurse should be able to perform highlight possibilities for better engineered future designs. mentally, but the nurse reported that the calculation was just The description of normal work provides a structure for not working for them at that point in time. The nurse making sense of the disturbances. proceeded by entering a trial value of 10ml for VTBI to go In this section, we consider three themes: the role of through to the time and rate settings. The nurse then entered observation in revealing such interaction issues; the role of one of these given values and saw what the pump calculated DiCoT in structuring the analysis; and possible for the remaining value. They could then see the calculated interventions to improve future designs. figure for the remaining value and deduce whether their Revealing invisible interaction issues guessed VTBI was higher or lower than that needed, and by Early discussions with the nurses indicated that there was what sort of margin. By performing this trial and error little wrong with the infusion pumps: they used the pumps workaround, the nurse worked out the correct VTBI. The frequently, they felt that they were well designed and they nurse then restarted the pump and programmed it correctly. did not have any interaction issues to report. However, Battery issue results reported here, in response to observational work The second issue we discuss is marked as disturbance 6 in rather than self-report, did find interaction issues. Figure 1: an infusion was manually stopped as soon as it We speculate that self-reporting failed because of the was started because the device had a low battery. Typically nurses’ “can-do” attitude in the face of problems; time all pumps are charged overnight on the Day Care Unit pressure; lack of vocabulary to articulate these HCI issues; ready for the next day. Pumps are run on their rechargeable and that they do not have the interest a HCI expert has in battery rather than being plugged in. One of the main these interaction issues. Interviews and questionnaires alone reasons for this is for mobility, both in terms of staff are limited for revealing these problems. moving them around the unit and the patients remaining mobile while receiving their treatment, e.g. so that they can As noted above, the issues discussed here have not featured go to the toilet. prominently in reported incidents that have, typically, resulted in serious harm. Reported battery life issues are A1 watched a nurse at intermittent times throughout the day more commonly associated with the poor retention of setting up successive parts of one patient’s treatment. The power, or battery failure, rather than cuing the user to nurse explained that some treatments last all day with a insufficient power at the point of programming. This design succession of different infusion programs. S/he remarked intervention has the potential to improve device and battery that you needed to be careful toward the end of the day management for nurses. Low battery power can be a because the device’s battery charge would not last for the problem when a socket is unavailable, e.g. when a patient is last treatment. S/he said that forgetting this was highly in transfer from one ward to another. In these situations the frustrating because you have to program a new pump to normally invisible interaction issue would become a finish the infusion with unfamiliar partial values. significant problem. Later that day, A1 was watching the nurse; s/he seemed to We note that clarifying the need for entering VTBI for the program everything correctly, pressed start, but then safe use of the pump has been remarkably difficult. It is immediately paused the pump. S/he pointed to the battery important to do this to understand the space for charge indicator, which was low, and said that it would not reengineering; however, the reasons for choosing VTBI as last. The nurse looked for a convenient socket to plug it in, the first value to be entered were not known by the clinical but then went to get a new pump that was fully charged and staff we had contact with, either on the day care unit or their reprogrammed the infusion with this new pump. 4 management team. In this sense, potentially important concern is the lack of clarity on what is possible and what is interaction design rationale is not known or visible. current practice, making definitive recommendations Due to their contextual nature, it is unlikely that these issues difficult: would have been discovered by analytic methods or Manufacturer laboratory studies alone. For example, it is recent advances In terms of the VTBI issue, the device’s instructions tell us in pump design that have introduced the battery issue: that the pump has been configured so the VTBI is a ‘target advances in technology have made infusion pumps small value’. This means that it must be entered first, then either enough to be easily mobile; older, larger pumps were the time or the rate, before the machine calculates the third. difficult to move around, and were therefore commonly If the second or third value is manipulated then the target stationary. Whilst stationary, their battery would only be value should remain the same whilst the corresponding third used for back-up, and so the battery issue would not have or second value is automatically adjusted; e.g., if the time is been a problem. changed then the VTBI should remain the same and the rate These two results were unremarkable disturbances in the should adjust accordingly. nurses’ normal work which, without observation, would Discussions with health services staff have revealed that the remain unreported, unnoticed and invisible. For the nurses device can be configured so that values can be entered in we observed having the difficulties, these are merely any order (this is the set-up in the intensive care unit). frustrations that could be alleviated. For the VTBI issue one However, devices in the Day Care Unit have been might need to use a bit more caution and mental effort to configured so that the user must enter the VTBI as the work out the VTBI manually. For the battery issue one ‘target value’. An untidy workaround to enter the time and might need to plug the infusion pump in to one of the many infusion rate so that the pump calculates the VTBI has been sockets around the unit, or programme a new pump partway developed by technical staff, but nurses do not know this, through an infusion. and it is far from ideal. However, we could imagine rare situations where these The battery issue is more clear-cut, in that this is a could contribute to an incident if unresolved. Indeed, the manufacturing design intervention, and not to do with local safety literature often refers to accidents as an unfortunate training, configuration, or management; i.e. the device combination of multiple minor failures rather than having a could be designed to warn the user if the programmed single main cause [6]. For example, imagine a novice nurse, treatment time will outlast the battery at the point of in an emergency, who is trying to work out the VTBI programming. During introductory meetings with the manually because s/he cannot skip this stage. At the same manufacturers of the observed pumps we raised this issue time another pump’s alarm disturbs her/him to signify it is and proposed this intervention; this suggestion was well running out of battery charge: s/he forgot to check the received. battery indicator when s/he programmed it. S/he switches Local Training, Configuration and Management attention to changing the second pump. In trying to In terms of the VTBI issue, some staff assert that all calculate the dose for the new pump s/he confuses it with prescriptions have the VTBI available, which contradicts the other VTBI calculation and enters too high an infusion other accounts. The nurse we observed understood that the rate; the patient comes to harm. This is only illustrative, but VTBI value was not available to her. We speculate that experience tells us to prepare for the unanticipated [6]. some doctors or pharmacists might not include this in their The role of DiCoT in the analysis handwritten prescriptions if they do not recognise the The process diagram shown above (Figure 1) is one of importance of doing so. If VTBI is always present, then many representations developed as part of this analysis. training should focus more on where the VTBI can be Others include representations of the device interface and found; otherwise, training needs more focus on how to of spatial layouts. As others (e.g. [14]) have noted, the quickly and reliably calculate VTBI from time and rate. details of healthcare work are messy, and it is essential to Alternatively, management might review policies and have an appropriate structuring representation to guide procedures. For example, if not entering the VTBI first observations, and to organize information to support does not pose any risk to patient safety then the pumps sensemaking. DiCoT served such a role in this study. could be configured so that any value can be entered, which Without such a structuring representation to focus data is the set-up in the intensive care unit. Alternatively the gathering and analysis, the task might have become policy would need to state that there is an accurate VTBI intractable. for every prescription. Socio-technical intervention CONCLUSION Ideally, we would like to make interventions to alleviate In this position paper, we have discussed the roles of interaction issues. We discuss different socio-technical observation and analysis structured around Distributed interventions in response to our results below; this work is Cognition in informing the engineering of medical devices on-going, so we report it as work-in-progress. An important that are better suited to their intended context of use. This 5 work is at an early stage of development; for example, it is 8. Hutchins, E. (1995) Cognition In The Wild. MIT Press, essential to conduct similar studies in different wards, in Cambridge, MA. different hospitals, and with devices from different 9. Hutchins, E. (1995) How a Cockpit Remembers Its manufacturers. However, this study has illustrated the value Speed. Cognitive Science, 19, 265-288 of DiCoT as a framework for structuring data gathering and analysis, and has also highlighted the importance of 10.ISMP Canada (2007) Fluorouracil Incident Root Cause conducting observational studies of normal work, and of Analysis. Available from http://www.cancerboard.ab.ca/ not relying on incident reports or self-report as the principal NR/rdonlyres/4107CCF0-2608-4E4D-AC75- data sources for informing future design decisions. E4E812F94FD6/0/Incident_Report_UE.pdf. 11.Lin, L., Isla, R., Doniz, K., Harkness, H., Vicente, K. & ACKNOWLEDGMENTS We thank the staff for their cooperation and patience with Doyle, D. Applying Human Factors to the Design of the study. This work was funded by EPSRC Grant Medical Equipment. J Clin Monit, 14, 253-263. 1998. EP/G059063/1. 12.MAUDE: http://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfma REFERENCES ude/search.cfm 1. Carayon, P., Hundt, A. & Wetterneck, T. (2010) Nurses’ Acceptance of Smart IV Pump Technology. Int. J. 13.Nemeth, C., Cook, R., O’Connor, M. and Klock, P. A. Medical Informatics, 79(6), 401-411. (2004) Using cognitive artifacts to understand distributed cognition. IEEE Transactions on Systems, 2. Carayon, P., Wetterneck, T., Hundt, A., Ozkaynak, M., Man and Cybernetics, Part A: Systems and Humans. Ram, P., DeSilvey, J., Hicks, B., Robert, T., Enloe, M., 34.6. 726-735. Sheth, R. & Sobande, S. (2005) Observing Nurse Interaction with Infusion Pump Technologies. Advances 14.Nemeth, C.P. Cook, R.I. Woods, D.D. (2004) The in Patient Safety: Vol 2. AHRQ. Messy Details: Insights from the Study of Technical Work in Healthcare. IEEE Transactions on Systems, 3. Furniss, D. & Blandford, A. (2006) Understanding Man and Cybernetics, Part A: Systems and Humans. Emergency Medical Dispatch in terms of Distributed 34.6. 689 – 692. Cognition: a case study. Ergonomics, 49(12-13), 1174- 1203. 15.Obradovich, J. & Woods, D. (1996) Users as Designers: How people cope with poor HCI design in computer- 4. Furniss, D. & Blandford, A. (2010) DiCoT Modeling: based medical devices. Human Factors, 38(4), 574-592. From Analysis to Design. Proc. CHI 2010 Workshop on Bridging the Gap: Moving from Contextual Analysis to 16.Scottish Executive (2006) Unintended overexposure of Design. patient Lisa Norris during radiotherapy treatment at the Beatson Oncology Centre, Glasgow in January 2006. 5. Hollan, J.D., Hutchins, E.L. & Kirsh, D. (2000) Available from www.scotland.gov.uk/Resource/Doc/ Distributed cognition: toward a new foundation for 153082/0041158.pdf. human-computer interaction research. ACM Transactions on CHI, 7.2, 174-196. 17.Wright, P.C., Fields, R.E. & Harrison, M.D. (2000) Analysing Human–Computer Interaction as Distributed 6. Hollnagel, E. Barriers and Accident Prevention. Cognition: the Resources Model. Human–Computer Ashgate. 2004. Interaction Journal. 15. 1-41. 7. Husch, M., Sullivan, C., Rooney, D., Barnard, C., Fotis, 18.Xiao, Y. (2005) Artifacts and collaborative work in M., Clarke, J. & Noskin, G. (2005). Insights from the healthcare: methodological, theoretical, and sharp end of intravenous medication errors: implications technological implications of the tangible, Journal of for infusion pump technology. In Quality and Safety in Biomedical Informatics, 38.1, 26-33. Health Care, 14: 80-86 6