Enhancing the interactivity of a clinical decision support system by using knowledge engineering and natural language processing

Abstract

Mental illness is a serious health problem and it affects many people. Increasingly,Clinical Decision Support Systems (CDSS) are being used for diagnosis and it is important to improve the reliability and performance of these systems. Missing a potential clue or a wrong diagnosis can have a detrimental effect on the patient's quality of life and could lead to a fatal outcome. The context of this research is the Galatean Risk and Safety Tool (GRiST), a mental-health-risk assessment system. Previous research has shown that success of a CDSS depends on its ease of use, reliability and interactivity. This research addresses these concerns for the GRiST by deploying data mining techniques. Clinical narratives and numerical data have both been analysed for this purpose.Clinical narratives have been processed by natural language processing (NLP)technology to extract knowledge from them. SNOMED-CT was used as a reference ontology and the performance of the different extraction algorithms have been compared. A new Ensemble Concept Mining (ECM) method has been proposed, which may eliminate the need for domain specific phrase annotation requirements. Word embedding has been used to filter phrases semantically and to build a semantic representation of each of the GRiST ontology nodes.The Chi-square and FP-growth methods have been used to find relationships between GRiST ontology nodes. Interesting patterns have been found that could be used to provide real-time feedback to clinicians. Information gain has been used efficaciously to explain the differences between the clinicians and the consensus risk. A new risk management strategy has been explored by analysing repeat assessments. A few novel methods have been proposed to perform automatic background analysis of the patient data and improve the interactivity and reliability of GRiST and similar systems.