A System for Conversational Case-Based Reasoning in Multiple-Disease Medical Diagnosis
In this thesis, we develop a model that uses Conversational Case-Based Reasoning (CCBR) in order to help physicians diagnose patients. To be able to process the vast amount of information embedded in the domain of general medicine, we introduce a divide and conquer approach. By focusing on small, well-defined sub-domains of medicine, we are able to capture specific knowledge from each of them. Together the sub-domains form our understanding of the medical domain, and we argue that this approach is more sound than to reason from the entire domain at the same time.
We adopt a set of existing approaches to the CCBR process to fit our needs. By testing these algorithms on real life data and analysing the results, we are able to identify strengths and weaknesses for each of them. By studying different dialogue management techniques embedded in CCBR, we are able to introduce targeted measures to increase the performance of these algorithms. At the same time, we are able to increase their flexibility, enabling them to take on domains that they previously did not support. We also introduce different dialogue inference techniques to our system, and demonstrate that this has the potential to further increase the performance of our system.
To bind the different sub-domains together we introduce an architecture that includes a stack of CCBR dialogues. This enables our system to explore multiple areas of medicine within the same session, increasing the probability of finding the correct diagnosis. For each sub-domain the system can choose from the set of CCBR algorithms included in the system, and find the one that maximises the performance in that particular domain. To be able to determine which dialogues to add to this stack we introduce a meta-level dialogue. This dialogue is added on top of the other dialogues and presents the user with a set of general questions in an effort to identify the most relevant sub-domains to explore.
Place, publisher, year, edition, pages
Institutt for datateknikk og informasjonsvitenskap , 2014. , 134 p.
IdentifiersURN: urn:nbn:no:ntnu:diva-26796Local ID: ntnudaim:11649OAI: oai:DiVA.org:ntnu-26796DiVA: diva2:751704
Aamodt, Agnar, Professor