Information Extraction from Patients' Free Form Documentation

The paper presents two rule-based information extraction (IE) from two types of patients’ documentation in Polish. For both document types, values of sets of attributes were assigned using specially designed grammars. 1 Method/General Assumptions Various rule-based, statistical, and machine learning methods have been developed for the purpose of information extraction. Unfortunately, they have rarely been tested on Polish texts, whose rich inflectional morphology and relatively free word order is challenging. Here, we present results of two experiments aimed at extracting information from mammography reports and hospital records of diabetic patients.1 Since there are no annotated corpora of Polish medical text which can be used in supervised statistical methods, and we do not have enough data for weakly supervised methods, we chose the rule-based extraction schema. The processing procedure in both experiments consisted of four stages: text preprocessing, application of IE rules based on the morphological information and domain lexicons, postprocessing (data cleaning and structuring), and conversion into a relational database. Preprocessing included format unification, data anonymization, and (for mammography reports) automatic spelling correction. The extraction rules were defined as grammars of the SProUT system, (Drożdżyński et al., 2004). This work was partially financed by the Polish national project number 3 T11C 007 27. SProUT consists of a set of processing components for basic linguistic operations, including tokenization, sentence splitting, morphological analysis (for Polish we use Morfeusz (Woliński, 2006)) and gazetteer lookup. The SproUT components are combined into a pipeline that generates typed feature structures (TFS), on which rules in the form of regular expressions with unification can operate. Small specialized lexicons containing both morphological and semantic (concept names) information have been created for both document types. Extracted attribute values are stored in a relational database.2 Before that, mammography reports results undergo additional postprocessing — grouping together of extracted data. Specially designed scripts put limits that separate descriptions of anatomical changes, tissue structure, and diagnosis. More details about mammography IE system can be found in (Mykowiecka et al., 2005).