Integrated representation of clinical data and medical knowledge: an ontology-based approach for the radiology domain

The increase in available clinical data and availability of medical knowledge provides the basis for various applications which aim to increase the quality and safety of healthcare. Personalized treatment, e.g. the optimal medication with fewest side effects, less unnecessary examinations and mistreatment can be realized based on the available data. Today however enhanced quality of treatment on the one hand and lower cost on the other hand are in conflict: in the current situation more data means more efforts for the clinicians and thus higher costs. This is mostly because of three problems: Firstly, clinical data is stored case-centric and distributed across different systems. It is not longitudinal integrated. Secondly, only small amount of the data is structured while high percentage of clinically relevant data is unstructured. Thirdly, existing data is not sufficiently aligned to medical knowledge and thus not on the appropriate level of detail for decision support systems. As a result of these problems most of the available data is simply not used in their full strength and no personalized treatment is applied. Today healthcare providers do not achieve improvements in quality and efficiency. The thesis has three objectives targeting parts of the aforementioned problems. The first objective is the creation of a semantic model for clinical information which is based on established upper ontologies. In particular, it is focused on the representation of clinical findings from radiology examinations. The second objective is to extract structured representations from radiology reports using formalized medical knowledge. The extracted information is stored in the semantic model. The third objective is to enrich the data using inference techniques and formalized medical knowledge to allow realization of different views on the data, needed by clinicians for more efficient decision making. In particular, radiology findings extracted from unstructured reports are classified as normal/abnormal and finding descriptions are linked to disease information. A longitudinal view of radiology finding data is realized through a prototype implementation of a report viewer which relies on medical knowledge about the human anatomy, the semantic representation of finding descriptions and the meta-data of the original radiology reports