Cognitive processing in a medical diagnosis task

Peter Yule

Approaches to Judgement and Decision Making are frequently driven by mathematical/normative theories. Such approaches rarely address important questions concerning processing. In order to address such questions we have investigated human behaviour on a task with a strong sequential component. Subjects take the role of a doctor attempting to diagnose a series of patients. Subjects are presented with an initial symptom, and allowed to query a set of further symptoms. At any point during a trial they are able to offer a diagnosis. Feedback on this diagnosis allows them to learn symptom/disease associations. Symptoms and diseases are associated probabilistically, and data from two conditions, differing in the number of symptoms associated with each disease, have been collected.

Subjects' learning within the task is evidenced by decreased questioning and increased diagnostic accuracy, but performance is dependent upon the number of symptoms associated with each disease. Subjects also show questioning preferences: their first query is influenced by the presenting symptom. In order to relate these data to processing considerations we have explored three cognitive models of subject behaviour (including learning) on the task: a Bayesian model, a symbolic hypothesis generating model, and an Associationist model. All models learn to perform at high levels, typically better than the human sample, and diagnostic accuracy improves as the task progresses. Symptom query rates tend to decrease as the task progresses, in both conditions, for the Bayesian and Associationist models, but the Symbolic model shows a qualitative difference in processing between these conditions, arguably corresponding to the difference between conditions seen in subjects.

We also compare model and human questioning strategies. The Bayesian model, which selects the most informative available symptom first, exhibits little similarity with the human performance, whereas the Symbolic model, which looks for symptoms positively associated with any hypothesised disease, shows the greatest fit with the human data. The Associationist model fits the human data to a middling extent. The good fit of the Symbolic model is contingent on assumptions about memory access as well as questioning strategy; this finding constitutes a non-trivial replication of the results of Fox (1980) concerning non-learning versions of these models.

Cognitive representation and processing of uncertainty: a mental models approach Workshop 3 The Young & O'Shea subtraction model