Knowledge extraction and analysis of medical text with particular emphasis on medical guidelines

In this dissertation document, we describe the potential for Information Extraction, Information Retrieval, and Machine Learning methods to improve the process of analyzing medical texts and, in particular, Clinical Practice Guidelines (CPGs). We present the results of three in-depth studies consisting of dozens of experiments on finding condition-action and other conditional sentences in guideline documents. We are improving the state-of-the-art results (from 5% to 17%) and showing for the first time the applicability of domain adaptation and transfer learning to this problem.
We also present new methods for identifying inconsistencies in disagreements between medical guidelines, and for analyzing them using a combination of machine learning, information retrieval, and text mining methods. We show the need for a formal distinction between contradictions and disagreements in natural language texts to formally reason between contradictory medical guidelines.
We introduce new representations for collections of guideline documents and an algorithm for comparing collections of documents. We use these to investigate conceptual distances between guidelines for the same conditions. Throughout this process, we prove the hypothesis that the difference in recommendations largely (by 69% to 86%) correlates with the differences in concepts used by the medical bodies authoring the guidelines.
Finally, we show the applicability of text analysis methods to practical problems of analyzing textual information in electronic health records. We achieved 83% accuracy in matching medical records with a list of pre-defined conditions in an EHR system, resulting in clinical system support changes in one of the leading US hospitals.