Unified Extraction of Health Condition Descriptions

This paper discusses a method for identifying diabetes symptoms and conditions in free text electronic health records in Bulgarian. The main challenge is to automatically recognise phrases and paraphrases for which no ”canonical forms” exist in any dictionary. The focus is on extracting blood sugar level and body weight change which are some of the dominant factors when diagnosing diabetes. A combined machine-learning and rule-based approach is applied. The experiment is performed on 2031 sentences of diabetes case history. The F-measure varies between 60 and 96% in the separate processing phases.