Supporting Healthcare Decision Modelling with Natural Language Processing

Medical and health policy decision makers attempt to answer complex and difficult questions with significant societal and economic implications, e.g. What are the benefits of minimum alcohol pricing? What is the cost effectiveness of screening newborns for X-ALD? Healthcare decision modelling answers these questions by identifying, assessing and synthesising relevant evidence to create models that can be used to inform decision makers. However, creating these models is a complex problem that is made difficult by the amount of potentially relevant information available. Natural Language Processing offers a solution that has potential to assist with the organisation and analysis of the evidence.
The project will develop novel Natural Language Processing techniques that support modellers in making sense of the data available and deriving models from it. It will develop and assess the effectiveness of technologies in three main areas:
(1) Automatic literature organisation. Healthcare decision models require evidence on a broad range of factors including disease epidemiology, natural history and management, comparative effectiveness of existing and novel interventions, costs, resources and health impacts. Current search techniques either rely on using broad search criteria or sets of highly focused searches. Statistical topic models will be applied to cluster documents containing evidence to generate an overall picture of the major factors that influence the model.
(2) Identification of key concepts and relations. Health economic models bring together complex networks of evidence to make predictions about the potential health and cost impact of novel interventions. The design of these models involves identification of sets of concepts and relationships between concepts. Interpretation of a large and diverse evidence base can be time consuming and difficult to replicate, tools to aid in this process have the potential to aid efficiency and effectiveness. Named entity identification will be applied to determine the main concepts discussed within the documents containing relevant evidence. Relation extraction will be applied to determine relationships between the concepts (for example "X causes Y").
(3) Model induction. The output from (2) will be analysed to derive directed graphs in the form of cognitive or causal maps that link interventions to outcomes and goals. Analytical routines will be developed to aid analysis of these maps which explore a) filtering of relations based upon weight of evidence and/or centrality of concepts to goals, b) identifying negative or positive feedback loops within graphs, c) visualisation of graphs for topic expert validation, d) generation of Markov models from graphs that form first approximations of healthcare decision models. Graph-based algorithms will be used in the definition of analytical routines.