Presented at the 1999 APIII Conference Return to 1999 Abstract Index
CREATING A DIAGNOSTIC TRACE: A METHODOLOGY TO STUDY VISUAL SEARCH, KNOWLEDGE REPRESENTATION AND MEDICAL REASONING IN MICROSCOPIC DIAGNOSIS
University of Pittsburgh
Center for Biomedical Informatics
Pittsburgh, Pennsylvania
Rebecca Crowley,
MD
The act of microscopic diagnosis seems mysterious. Somehow
the combination of a trained eye, a microscope, and H&E
slide produce a classification of disease – an act
of paramount clinical importance. But ask a pathologist
"How do you recognize that as a case of nodular fasciitis?"
and they are likely to respond "It just looks like
one!". More introspective pathologists may provide
a list of criteria: "It’s composed of spindle
cells with a tissue culture appearance". But few will
be able to tell you exactly how they searched the slide,
selected the areas of diagnostic importance, distilled the
mass of visual information into a set of diagnostic features,
generated a differential diagnosis, and deduced the correct
entity.
Drawing our theoretical framework from the field of human information processing, we are developing and testing a methodology to study these processes. Our findings will form the basis for a model of diagnostic pathology expertise. In future work, we hope to use this model in developing expert systems in pathology, including decision support systems and intelligent tutors.
Our novel methodology combines analysis of videotaped microscope feed, verbal "think aloud" protocol analysis, and automated methods for capturing stage location and magnification over time. Using these techniques we are able to generate:
- visual representations of all stage movements and magnification changes made as a function of time
- verbatim transcripts of pathologists reporting exactly what they are thinking as they look at a slide
- a permanent videotaped record of the microscopic fields as they are viewed.
Data from videotapes and verbatim transcripts are coded and analyzed, and then combined to produce a wide variety of timed and counted measurements. In addition, verbal protocols are used to generate semantic networks that represent the knowledge and reasoning components of this task. By combining all of these data representations, we create a trace of the diagnostic process itself.
Using these techniques, we have completed an exploratory study of a small number of expert (attending) and novice (resident) pathologists. Preliminary findings show substantial differences between experts and novices in terms of diagnostic search patterns, knowledge representations and reasoning.