519 Distinguishing Existing Findings from Planned Treatments in Dictated Dental Exams

Thursday, March 4, 2010: 3:30 p.m. - 4:45 p.m.
Location: Exhibit Hall D (Walter E. Washington Convention Center)
Presentation Type: Poster Session
D. MOWERY, T. SCHLEYER, and W. CHAPMAN, University of Pittsburgh, Pittsburgh, PA
Objective: Understanding the context of information extracted from medical text is important for biomedical applications that facilitate workflow in patient care settings and monitor quality of care. A quality assurance application needs to assess whether a clinician has adhered to treatment guidelines for tobacco cessation by arranging follow-up care plans (e.g. schedule a visit to review progress towards quitting). A dental charting application that identifies findings needs to discern existing conditions (e.g. crown on 14) from planned conditions (e.g. needs crown on 14). The objective of this project is to build a classifier for the medical and dental domains that determines whether a finding occurs in the context of a treatment plan. Methods: We applied machine learning algorithms to words in emergency department dictations to develop a medical-plan classifier that classifies sentences as plan or not plan. We created the classifier using 10 emergency department reports and tested it on 10 reports. Results: Words like “will” and “as needed” were good predictors for classifying sentences as a plan. Error analysis of both Ripper (Recall: 47.8%, Precision: 68.8%) and Decision Tree (Recall: 23.9%, Precision: 61.1%) indicate additional features that should improve prediction. Conclusion: Semantic mapping (e.g. mapping conditions from the text to UMLS concepts) and contextual features (e.g. where in the report the sentence is retrieved) may provide more information gain and increased system performance. Using the rules learned by the medical-plan classifiers, we are developing a rule-based dental-plan classifier. We intend to analyze the differences between performance on medical reports and dental exams. This research is supported by the NLM/NIDR Fellowship: 5T15LM007059.

Keywords: Behavioral science, Computers, Dental Informatics, Interfaces and Technology