Claudia Bretschneider

Towards a technology roadmap for big data applications in the healthcare domain

Big Data technologies can be used to improve the quality and efficiency of healthcare delivery. The highest impact of Big Data applications is expected when data from various healthcare areas, such as clinical, administrative, financial, or outcome data, can be integrated. However, as of today, the seamless access to the various healthcare data pools is only possible in a very constrained and limited manner. For enabling the seamless access several technical requirements, such as data digitalization, semantic annotation, data sharing, data privacy and security as well as data quality need to be addressed. In this paper, we introduce a detailed analysis of these technical requirements and show how the results of our analysis lead towards a technical roadmap for Big Data in the healthcare domain.

UIMA2LOD: Integrating UIMA Text Annotations into the Linked Open Data Cloud

The LOD cloud is becoming the de-facto standard for sharing and connecting pieces of data, information and knowledge on the Web. As of today, means for the seamless integration of structured data into the LOD cloud are available. However, algorithms for integrating information enclosed in unstructured text sources are missing. In order to foster the (re)use of the high percentage of unstructured text, automatic means for the integration of their content are needed. We address this issue by proposing an approach for conceptual representation of textual annotations which distinguishes linguistic from semantic annotations and their integration. Additionally, we implement a generic UIMA pipeline that automatically creates a LOD graph from texts that (1) implements the proposed conceptual representation, (2) extracts semantically classified entities, (3) links to existing LOD datasets and (4) generates RDF graphs from the extracted information. We show the application and benefits of the approach in a case study on a medical corpus.

Corpus-based Translation of Ontologies for Improved Multilingual Semantic Annotation

Ontologies have proven to be useful to enhance NLP-based applications such as information extraction. In the biomedical domain rich ontologies are available and used for semantic annotation of texts. However, most of them have either no or only few non-English concept labels and cannot be used to annotate non-English texts. Since translations need expert review, a full translation of large ontologies is often not feasible. For semantic annotation purpose, we propose to use the corpus to be annotated to identify high occurrence terms and their translations to extend respective ontology concepts. Using our approach, the translation of a subset of ontology concepts is sufficient to significantly enhance annotation coverage. For evaluation, we automatically translated RadLex ontology concepts from English into German. We show that by translating a rather small set of concepts (in our case 433), which were identified by corpus analysis, we are able to enhance the amount of annotated words from 27.36 % to 42.65 %.

Semantic Splitting of German Medical Compounds

Compounding is widespread in highly inflectional languages with a quarter of all nouns created by composition. In our field of study, the German medical language, the amount of compounds significantly outnumbers this figure with 64i¾?%. Thus, their correct splitting is a high-impact preprocessing step for any NLP-based application. In this work we address two challenges of medical decomposition: First, we introduce the consideration of unknown constituents in order to split compounds that were not recognized as such so far. Second, our approach builds on the corpus-based approach of Koehn and Knight and adds semantic knowledge from domain ontologies to increase the accuracy during disambiguation of the various split options. Using this first-of-a-kind semantic approach in a study on decomposition of German medical compounds, we outperform the existing approaches by far.

Knowledge-Based Extraction of Measurement-Entity Relations from German Radiology Reports

A large percentage of relevant radio logic patient information is currently only available in unstructured formats such as free text reports. In particular measurements are important since they are comparable and thus provide insight into the change of the health status over time, for example in response to some treatment. In radiology most of the measurements in reports describe the size of anatomical entities. Even though it is possible to extract measurements and anatomical entities from text using standard information extraction techniques, it is difficult to extract the relation between the measurement and the corresponding anatomical entity. Here we present a knowledge-based approach to extract this relation for size measurements using a model about typical size descriptions of anatomical entities in combination with hierarchical knowledge of existing medical ontologies. We evaluate our approach on two data sets of German radiology reports reaching an F1-measure of 0.85 and 0.79 respectively.