Monday, July 04, 2005

Reading :: Rationalizing Medical Work

Originally posted: Mon, 04 Jul 2005 19:32:13

Rationalizing Medical Work: Decision Support Techniques and Medical Practices

by Marc Berg

Marc Berg's dissertation work was supervised in part by Annemarie Mol, and her influence is quite evident in this book based on the dissertation. Berg uses actor-network theory to examine medical work, particularly the trend of rationalization that is exemplified by decision support systems -- information systems that are meant to help doctors make consistent, accurate decisions about medical care. Such systems look good on paper, but begin to provide wildly inaccurate diagnoses when applied to problems outside their specialty area. Berg wants to know: Why is this so? How do medical personnel deal with the problem? What does all of this tell us about knowledge work?

As I've discussed elsewhere on this blog, Berg has done some very smart work in relating ANT to computerized information systems, and in doing so he has brought it onto "turf" that has been occupied by information processing cognitive psychology and (more recently) situated cognition, distributed cognition, and activity theory. ANT has some methodological and theoretical relations with the latter three, but it is ultimately a rather different enterprise because it represents a relationist sociology hooked up to a particular type of ontology. So although it deals with some of the same problems that an AT or DC study might, it sometimes diverges in interesting ways, especially as it treats history, multiplicity, and development.

Berg discusses how medicine has been framed over the last couple of centuries, noting that it has involved a struggle between two concepts of medicine: as a situated practice grounded in tacit knowledge and as a set of general principles grounded in formal research. The latter view has gained the upper hand lately, resulting in attempts to flatten and univeralize concepts and vocabularies (e.g., p.24). This view has led to developing automated decision systems that can compile general knowledge about various conditions, based on the research paradigm. Such systems not only draw from normative concepts and vocabularies, it enforces them: the practice of its users is "disciplined to this formalism" (p.92, his emphasis).

Unfortunately, this approach simply doesn't work well. Like Mol, Berg identifies the problem as multiplicity: different disciplines simply don't and can't agree on the "same" object (p.96). As one informant complains: "But nobody agrees on what they are talking about. What should the result of the test be. A figure? The angle the leg makes with the table?" The informant goes on to describe ever more specific variations of decisions physicians must make in order to reach the degree of specificity necessary for agreement. Berg's point is that such negotiations could go on in remarkable detail for each case, with more negotiations necessary whenever a new medical specialty is brought in, and no guarantee of agreement -- because the different specialties are looking at different objects and mobilizing them in different ways. The approach necessitated by a decision support system involves building these negotiations into the tool itself, flattening them into a single discipline with a single body of knowledge and object. The task is problematic to say the least:

... formal medical tools cannot be conceived ass inert carriers of some "good medical practice." Delegating tasks to a formal tool transforms the nature of those tasks. The introduction of a decision-support tool generates a propensity to refocus medical criteria on the elements that behave in predictable and easily traceable ways. Formal tools contain a predisposition to build simple, robust worlds, without too many interdependencies or weak spots where contingencies can leak back in. In doing so, in selecting the measurements and indications that best fit its prerequisites, the breast cancer protocol redefines what eligibility for bone-marrow transplantation treatment denotes -- and, thus, what "potentially curable disseminated breast cancer" is. (p.99)

Hard data, Berg says, is data whose production has been disciplined (p.101).

So on the one hand a decision-support system redefines that which it describes, but on the other hand the system cannot contain too many negotiations with too many disciplines. That quickly becomes too complex a task -- unless the system is sharply bounded. "Constructing a feasible, working decision-support tool, then, always implies building specific contexts into the technique" (p.108). The decision-support system appears to embody some sort of universal medical knowledge, but it can only do so in a sharply limited space, and even then, "idiosyncratic, unique features of the specific sites involved become embedded in a tool's script" and thus "a tool's radius of action is reduced" (p.108). "The contingent nature of the protracted process of negotiations, I argue, is incorporated into the core of the tool. Trying to get a protocol to work is a process of making ad hoc compromises, going back to the tool, and tinkering to get the medical practice's elements in line" (p.115). The traces of these struggles and compromises are left in the tool (p.116).

Ultimately, "Instead of the transparent, optimal, unified Clinical Rationality hoped for, we end up with opaque, impure, additional rationalities" (p.116).

In the latter part of the book, Berg brings the principle of symmetry to bear on the issue. He points out how chains of events are consistently rewritten to attribute decisions to individual physicians rather than to the tools and practices on which they draw, and in doing so, he calls into question the notion of uniquely human agency (p.136; cf. Hutchins, Suchman). Berg also follows ANT by arguing that history and future are continually reconstructed in medical records (p.137).

Ultimately, Berg argues, "the only way the network can persist is through its looseness, its openness, and its unresolved tensions" (p.168). I agree, of course, having made a similar argument in my own book.

Overall, Berg makes a strong case for applying ANT insights to information systems. Those of us coming from an activity theory standpoint can extrapolate a critique of our own work as well as ways in which the two approaches complement each other.

Blogged with Flock

No comments: