Processing clinical guideline text for formal verification

Abstract

Clinical guidelines are evidence-based recommendations developed to assist practitioners in their decisions on appropriate care for patients with specific clinical circumstances. They provide succinct instructions such as what drugs should be given or taken for a particular condition, how long such treatment should be given, what tests should be conducted, or other situational clinical circumstances for certain diseases. However, as they are described in natural language, they are prone to problems such as ambiguity and incompleteness. As the guidelines are publicly accessible, we expect them to be foolproof from inconsistencies and missing gaps. This thesis aims to answer a couple questions in regard to the correctness of clinical guidelines: (1) How can we get the main information in clinical guideline texts? (2) How can we check the guidelines in terms of correctness and consistencies? To answer these questions, first, we develop several methods to mark and capture the semantic information in the texts. We start by building a controlled natural language to reduce the complexity of the texts’ structure. We show that this approach is easy to set up but obviously unscalable. We then consider machine learning approaches and use semantic role labelling, named-entity recognition and relation classification techniques. To achieve this task, we create a clinical guideline corpus tagged with process labels. We show that even with a small corpus, the baseline performance is promising. We then investigate fine-tuning some state-of-the-art neural model architectures and get better performance. Finally, we create a framework to transform the clinical guidelines into formal statements and check their correctness against some properties using model checkers or constraint solvers. This thesis presents a study and analyses of entity labelling and relation classification in regard to clinical guidelines, as well as formally checking their correctness, providing insights and future research directions on the improvement of clinical guidelines.