Uncelebrated Case #2:
Antibody misidentification and transfusion reactions

Smith and colleagues studied the factors that contribute to antibody misidentification. The studies of expertise led to the development and testing of new systems to aid lab technician performance on this task and to support improved training. Tests of the new system have shown substantial decreases in antibody misidentification.

Background
There are a number of antigens that may be present on red blood cells. These can cause a transfusion reaction if a patient sensitive to these antigens receives blood that contains them. To avoid these reactions, lab technicians check to see which antibodies are present in a patient's blood prior to transfusion.

Blood screening checks for the presence of a variety of antibodies in the patient's blood using the results of a battery of tests performed in parallel. Taken together, the results of these tests indicate the types of antibodies present. The results are presented in a table that relates reactions to hypotheses about what antibodies may be present. The technicians evaluate the pattern of reactions and make inferences about which antibodies are present. Knowing which antibodies are present allows them to select blood units without corresponding antigens for transfusion. A crossmatch is then performed using the selected units of blood and the patient's own serum in order to verify the compatibility.

The research project was undertaken to develop improved means to train technicians for this task. In particular, the goal was to use new technology (e.g., artificial intelligence) and techniques (e.g., exploratory learning) to develop computerized tutor and learning aids.

The Research
The project was not driven by reactions to visible or celebrated failures, i.e., transfusion reactions leading to severe patient consequences. Instead, the motivation was a desire to demonstrate the use of new information technology and to assist the training of new technicians. The research team began with studies designed to examine the ways that experts perform the task and to explore the contrast between strategies used by experts and the behaviors of students and less experienced practitioners.

Multiple research methods were used, all focusing on (1) identifying what makes problems difficult and then (2) using difficult problems to understand what characterizes successful and poor problem-solving strategies. The methods used included critical incident analysis, knowledge elicitation based on walkthroughs of cases, observation of practitioners solving real problems in actual facilities, and observation of practitioners solving simulated problems using a high fidelity computerized test bed for exploring new approaches for support and training.

The research showed that failure occurs even with easy tasks because the tools people use create vulnerabilities. Antibody misidentification can arise from "slips" traced to characteristics of the paper tools used for record keeping. Given the identical rows and columns of the current paper format (Figure 4), it is relatively easy to start reading across a row and shift up or down so that the wrong row is scanned. This leads to a misinterpretation of the pattern of reactions. This link between the design of the paper forms and a form of failure suggests some straightforward interventions to improve performance. Thus the research points to some of the "low hanging fruit" that safety advocates believe should be exploited quickly to improve safety. It is worth noting, however, that being able to see this sort of opportunity is easy in hindsight, after the research is done-it was not recognized as sigificant before.

Some cases of antibody identification are difficult because they contain factors such as noisy data, weakly reacting antibodies, or multiple interacting antibodies that mask each other. Cases with these attributes often proved very difficult for both students and practicing lab technicians, and they frequently went "down the garden path." Misidentifications were surprisingly frequent for some cases with these characteristics with rates approaching 50% among practicing lab technicians.

Other problems are difficult to solve because they challenge the fund of knowledge that lab technicians possess. As in other areas studied (Feltovich, Ford and Hoffman, 1997), practitioners sometimes possess particular misconceptions that lead to poor performance on certain kinds of problems. For example, technicians sometimes had a misconception regarding the effects of pre-warming on reactions.

As these studies examined expert strategies, it quickly became apparent that more experienced practitioners had developed strategies that were sensitive to possibility of misidentification. Some were generic strategies that helped avoid traps or recover from a tentative misidentification. Others were tailored to help avoid specific vulnerabilities. Less expert practitioners did not possess these strategies for detecting and recovering from incipient misidentification.

Finding these kinds of results depended in part on knowing where to look. Past research on human performance on diagnostic tasks suggested that there would be classes of problems that offer subtle or infrequently encountered patterns that are more likely to lead to misidentifications. The investigators, as they learned more about which basic patterns were embodied in the antibody identification task, were able to predict where misidentifications would occur.

As the researchers began to understand what made certain kinds of problems difficult and how expert strategies were tailored to these demands, the researchers asked certain questions. What kinds of tools could be used to assist lab technicians? What kinds of training could improve their knowledge and strategies for difficult cases? What would help a lab technician recognize that a particular case was likely to be difficult? How could lab technicians decide when they needed help from more experience personnel to solve the case they faced?

Following these explorations, the simulation of antibody identification tasks became a test bed for exploring the impact of different strategies to improve performance. Some of these involved relatively simple perceptual and memory aids. For example, the electronic highlighting in Figure 4 helps the problem solver keep track of intermediate results. Other strategies involved having the computer do basic clerical housekeeping and tracking of tasks.

Since the research had uncovered expert strategies, especially ones that guard against or help recover from possible misidentifications, the designers developed a "critiquing" or advisory component in the computer system. An example of the output of this critiquing component is shown in the message in the lower window in Figure 4. In effect, the computer would "tap them on the shoulder" and say, "wait a minute, perhaps you should consider this before you go on: you just ruled out an antibody that seems inappropriate or you just left this panel without making some inferences that I (the computer) think are possible."

The researchers went on to explore different ways to use this "intelligent" capability. There are multiple ways to use this computer capability, and they are not all equally effective in improving the overall performance on the identification task. The overall performance of the human-computer team was significantly better when the human solved the identification problem and the computer provided a critique of the method used than when the computer solved the problem and the human provided the critique to ensure the computer had a correct solution. In both cases the knowledge in the computer was the same; the difference in performance was a result of the roles assigned to the people and to the computer.

This conclusion is consistent with other studies, namely that team performance is better when the computer plays the role of the critic. The difference can be large, in some cases as much as 30% better. This most probably is the result of a framing effect. This refers to the way that a suggestion from the computer can limit the variety of different possibilities that the human operator explores. When the computer suggests, practitioners tend to follow this proposed reasoning and agree with the computer even when the quality of the initial assessment by the computer is poor.

The timing and character of the computer critiques are also important in determining the overall performance of the human-computer team. In the end, the researchers were able to significantly reduce misidentification rates. According to their study, students who finish their formal curriculum and then spend 2 or 3 hours in this kind of learning environment improve their performance on test cases by about 90%. In a laboratory setting, with a critiquing system present, performance can improve between 30% and 60%.

The research project is complete. A system is currently available as stand-alone tutoring software for the cost of media. Several labs are using the software on their own. However, there are no formal technology transfer mechanisms in place, no active assessment programs to guide the transfer, and no software maintenance or support available. All these components necessary to translate research into improvements in safety are missing.

Implications of the Research
New technology is often proposed as the solution to a "human error" problem. This project illustrates that, while technology may well be part of system improvements, technology alone is not sufficient (Woods et al., 1994, chapter 5). In this case, the critical information to decide how to use technology skillfully came from:

• understanding what constitutes hard problems,
• understanding the ways in which the task of identification is vulnerable to failure, and
• understanding the strategies experts use to guard against and recover from trouble.

Note how the research results are not mere details of implementation or user acceptance that can be dealt with after the basic concept for new technology is implemented. The studies helped discover how to use technological possibilities to aid performance. The result was a cooperative concept that was quite different from the more autonomous machine that, in the absence of good data on the nature of expertise and failure, some expected to build.

The research strongly supports the use of decision-support tools to improve human performance. It also provides a warning about the limitations of automation as a replacement for human expertise. This case, with many others, provides an explanation for the repeated failures of efforts directed at replacing human expertise with machines. The results show that autonomous computer problem solvers are brittle, that people's judgment can be adversely influenced by the computer's behavior, and that a carefully constructed cooperative system that coordinates both human and machine expertise performs better than either one alone.

The work demonstrates quite clearly that creating effective computer-based decision aids is itself a complex task. It requires detailed knowledge of the ways in which human expertise is deployed, how it achieves success, and how it is vulnerable. There are few applications of computer aids to decision making in medicine that have been developed with such an understanding of the cognitive demands of practice in place. Ultimately, improvement in the overall system performance depends on increasing expertise, not replacing it. The design work revolved around developing mechanisms to enhance expertise. But expertise is already highly refined in complex work domains. Improving on it requires detailed understanding of the strengths and vulnerabilities of the current knowledge and strategies.

The value of the research is not limited to the development of a specific decision support tool. The knowledge necessary to produce the computer system can be used in many other ways. Discovering the components of expertise and making them explicit permit us to consider other ways to deliver enhanced expertise where it is needed. For example, the blood banking community uses case studies for testing expertise and performance in hematology labs. The knowledge base about difficulties, typical misconceptions, and expert strategies developed for the tutor system also could be used to enhance this process.

Expertise is not simply individual skill and knowledge about the narrow technical aspects of problems; it also refers to how an organization develops, supports, deploys, and brings to bear this narrow technical expertise in different kinds of situations.

In searching out the vulnerabilities in the current system for antibody identification, the research demonstrates that success depends at least as much on effective mechanisms for detection and recovery from incipient failure as it does on the primary prevention of flaws. The success of blood banking does not arise from the elimination of errors in the antibody screening process. Rather it relies on the recognition of cases that are special and prone to failure together with indications that the identification process has gone awry. The system does not perform flawlessly at each stage but rather manages to incorporate sufficiently sophisticated detection of flaws so that the overt failure rate is very low. One component of this is the expertise applied to screening; there are others.

This critical role of detecting and recovering from incipient failure is a fundamental finding of the new look at the human contribution to safety. This finding stands in stark contrast to the erroneous and overly simplistic notion that people are erratic and unreliable components in an otherwise successful system.

During the workshop a question was posed: if misidentification rates are so high for certain classes of problems (in cases with noisy data, weakly reacting antibodies, masking, misidentification rates approach 50% in data from practicing technologists), why aren't there more overt failures (transfusion reactions)? There are a number of reasons. The system is relatively tolerant of misidentification. The antibody identification process identifies only candidate blood units. Cross matching of the candidates with the patient's blood will detect many (but not all) incompatibilities. Even when there are incompatibilities, transfusion of the wrong unit may generate no significant problem, the problem may go unrecognized, or it may be attributed to some other source. The situation is, like so many others, complex. But an important consequence of the apparently low rate of mismatching as a source of injury is that we may expect to find little enthusiasm for expensive new programs to improve human or system performance. Paradoxically, the typical success of the blood testing system and the low-frequency of the sorts of complex identification problems used as test cases in this research may lead many to discount the value of such research.

Uncelebrated Case #3
Drug misadministrations via computerized infusion devices in the operating room
 

Contrasting Uncelebrated and Celebrated Cases

Taken together, the laparoscopic cholecystectomy, blood antibody identification, and infusion device cases demonstrate the kinds of insights that come from exploring the second story that lies behind the incidents that provoke attention. In each case, the work is painstaking and detailed, going far beyond the sorts of investigations that followed the celebrated cases. In each case the story is complex, difficult for outsiders to understand, and not easily reduced to a simple summary. Significantly, the research methods used are unfamiliar to many. Finally, the motivation for the work was less the desire to directly generate safety improvements than to understand the nature of the real processes that underlie success and failure in the real world. The potential for such work to produce sustained increases in safety is substantial. In particular, in each case the research offers the possibility of further progress by identifying areas ripe for additional work.

Day One Footnotes

See Obradovich, Smith, Guerlain, Rudmann, Strohm, Smith, Svirbely, & Sachs (1996) and Guerlain, Smith, Obradovich, Rudmann, Strohm, Smith, & Svirbely (1996).

Return to document

A generic class of problems have been termed "garden path" in research on problem solving because the pattern of initial evidence makes it easy for people to focus and become fixated on a plausible but erroneous diagnosis (see Woods et al., 1994).

Return to document

Note that the cooperative system is larger than just the human-machine team. It can also include multiple people, techniques to catch misidentifications, and certification processes, among other elements.

Return to document

One part of the original research team was interested in expert performance in order to develop a machine that could perform the task automatically. Interestingly, it turned out that cases that were hard for people were also hard for an artificial intelligence software system. As a result, the project emphasized using technology to support and cooperate with people.

Return to document

Problem solvers are brittle when they have narrow scope of competence and are unable to cope with problems that fall outside of that narrow scope. Research on autonomous machine problem solvers consistently finds that such machines are brittle. People, on the other hand, can be effective at adapting plans to handle complicating factors, surprising variations, and novel combinations. In other words, human and machine problem solvers are vulnerable to different kinds of failure.

Return to document