Formal Verification of CNL Health Recommendations

Abstract

Abstract. Clinical texts, such as therapy algorithms, are often described in natural language and may include hidden inconsistencies, gaps and potential deadlocks. In this paper, we propose an approach to identify such problems with formal verification. From each sentence in the therapy algorithm we automatically generate a parse tree and derive case frames. From the case frames we construct a state-based representation (in our case a timed automaton) and use a model checker (here UPPAAL) to verify the model. Throughout the paper we use an example of the algorithm for blood glucose lowering therapy in adults with type 2 diabetes to illustrate our approach.

Semantic Annotations in Clinical Guidelines

“…The main aim for this work is to help people retrieve the key information in clinical guidelines. As shown in [19], we built a system to do formal verification of a therapy algorithm for type 2 diabetes. As guidelines are expressed in natural language, they are prone to ambiguity, incompleteness, and inconsistency.…”

“…In future work, our aim is to build a framework that integrates the whole process defined in [19]. It means that we will also add functionalities to produce some formal models from the annotated guideline sentences.…”

“…In this research, we modify our previous work [19] following the approach used by Diamantopoulos et. al.…”

“…Knowing these semantic markers means that we can move on to transforming the guideline into any further modelling approach we have in mind, for example, a formal representation. We believe it would be beneficial to help the transformation process from guidelines to UPPAAL models as done in [19]. In order to build a learning model to mark the roles of a word or a phrase in a sentence, we first need to define the classes of roles that we allow in our domain.…”

“…We have built a framework to formalise the written text of a therapy algorithm for lowering blood glucose for people with type 2 diabetes (T2D) in our previous work [19]. The generated model is a network of timed automata, where each transition represents the medication taken by the patient and the state is the patient's condition after getting such treatment and over time.…”