Augmenting expertise

Cognitive studies of clinical comprehension reveal that expert clinicians are distinguished by their superior ability to recognize meaningful patterns of data in clinical narratives. For example, in psychiatry, the findings of hallucinations and delusions suggest the subdiagnostic hypothesis of a psychotic episode, which in turn suggests several diagnoses, including schizophrenia. This dissertation describes the design and evaluation of a system that aims to simulate an important aspect of expert comprehension: the ability to recognize clusters of findings that support sub-diagnostic hypotheses. The broad range of content in psychiatric narrative presents a formidable barrier to achieving this goal, as it contains general concepts and descriptions of the subjective experience of psychiatric patients in addition to general medical and psychiatric concepts. Lexically driven language processing of such narrative would require the exhaustive predefinition of every concept likely to be encountered. In contrast, Latent Semantic Analysis (LSA) is a corpus-based statistical model of language that learns human-like estimates of the similarity between concepts from a text corpus. LSA is adapted to create trainable models of sub-diagnostic hypotheses, which are then used to recognize related elements in psychiatric discharge summary text. The system is evaluated against an independently annotated set of psychiatric discharge summaries. System-rater agreement approached rater-rater agreement, providing support for the practical application of vector-based models of meaning in domains with broad conceptual territory. Other applications and implications are discussed, including the presentation of a prototype user interface designed to enhance novice comprehension of psychiatric discourse.