Comparison of treatment plan models

To assist physicians with the treatment planning process so-called clinical practice guidelines are created. They contain general information about a specific clinical condition as well as rules and procedures to treat patients with this condition. As these guidelines are represented in free text form they are difficult to handle for physicians in their daily working process. Therefore the information in the guideline document is formalized to create computerized guidelines. We work with the formalization language Asbru. The creation of these computer-interpretable guidelines is either done manually which is a great effort or automatically using methods of Information Extraction. The aim of this thesis is to verify whether the automatically generated model of a guideline corresponds to the manually generated model of the same guideline. Doing this manually would again be a great effort and therefore I want to investigate methods to automatically evaluate parts of these models. I focus on the procedural knowledge of the models. To be able to compare two different models I investigated them. At first I developed methods to compare two models. To make a statement about the similarity of two models we started with comparing their activities using similarity metrics in order to identify corresponding plans of two models. In a second step we furthermore looked into their process structure using workflow patterns and tried to find similarities as well. Then I implemented these methods prototypically and tested them using a `real-world' example, a guideline for gestational diabetes mellitus. The example was based on the output of the GESHER tool for manually creating computerized treatment plan models and the tool LASSIE for automatically creating computerized treatment plan models. Using this example we evaluated how much of the original guideline text is present in the automatically generated model. Finally, it was evaluated if the automatically generated model finds the same information as present in the manually generated model.