Human and Computer Attempts to Decode Symptom Language

Abstract

 

Clinicians are able to identify psychosomatic disease on the basis of interview transcripts, yet little is known about the clues they use to make their decisions. In an attempt to answer this question, segments from the therapy protocol of a former ulcer patient were assessed by clinicians and computer to detect discriminating signs. The words up and down were found (by a clinician) to be the two best single predictors of symptom frequency, accounting for 60% of the total variance. They appear to be redundant additions to the manifest content which are systematically inserted into the thought stream by the pressure of dynamic forces. Other marker words, which were embedded in idiomatic contexts, were found by computer. Because they are not crucial to understanding the manifest content, these clues can more easily bypass the self-scrutiny of the patient. But because they are redundant additions, they can only be detected if the therapist listens to words and not to meaning.