School of
Information Technology and Electrical Engineering

Speaker: Anton van der Vegt
Seminar Date: Tue, 15/10/2019 - 11:00
Venue: 49-313A; Advanced Engineering Building
Host: Dr Guido Zuccon

Seminar Type:  PhD Thesis Review

Abstract: 

Every day, doctors make clinical decisions that impact the health and longevity of patients, and more broadly, these same decisions, impact the effectiveness and efficiency of the health organisations they work within. Each clinical decision that a doctor makes is reliant upon both their innate knowledge and their ability to acquire the additional knowledge necessary to make a correct decision. Overlaid onto this decision making process is a challenging clinical environment that limits the time available to seek this additional knowledge. It is to the time and knowledge limitations of the clinician that this thesis directs its investigations; specifically, how minimal interaction information retrieval can be brought to bear on these limitations to enable more effective clinical decision making.

Clinicians are typically very busy and often work under considerable time pressure. Although medical search systems, such as PubMed, are widely available to support clinical decision making, clinicians report that lack of time is one of the biggest barriers to using such systems. In this thesis, a user study is conducted to clarify the effectiveness of document retrieval systems, such as PubMed, for time pressured clinicians. Our findings show that when clinicians are put under increased levels of time pressure, the benefit of the document retrieval system is significantly, and mostly, eroded. Further, clinicians only answered around half of the clinical questions correctly, and this was primarily because they were unable to interpret the documents correctly.

A novel task-based conceptual Information Retrieval (IR) framework is derived to support user search tasks that demand faster resolution with less document interpretation. It is based on the unit of Bridging Information, which is the additional information a user requires, beyond their innate knowledge, to complete their task. The framework supports an alternative approach to current document IR methods, called minimal IR (mini-IIR). Mini-IIR reduces the interaction required by returning a single information object, containing, for example, synthesized information or answers, rather than returning a list of documents.

A method is developed to utilise a mini-IIR approach to help clinicians to correctly answer their clinical questions. The method incorporates a new disorder centric, cross-sentence, relationship extraction component that utilises an LSTM deep learning model. This cross-sentence approach outperforms a range of benchmark systems. Turning to overall answer retrieval; the best performing mini-IIR system could find correct answers for 9 in 10 diagnosis-based clinical questions and could rank the correct answer in the top 10 for 4 in 10 questions.  One of the key challenges identified was relationship redundancy that is found in the evidential database, which biases the rankings.

The effectiveness of the mini-IIR system shows considerable promise for providing a much faster method for clinicians to identify answers for some of the questions confronting them, without having to search for and interpret complex medical literature. Extensive opportunities exist to improve the effectiveness of the mini-IIR system and these are detailed.

This thesis represents the first steps into a new line of research, minimal interaction IR, which is timely, as many more search tasks are being performed via voice or on small mobile screens, where interaction is limited. Our research provides the theoretical underpinnings for applications in clinical decision support, where we have demonstrated that mini-IIR methods are both possible and effective for some clinical questions. We are cautiously optimistic that mini-IIR will become an integrated element of future clinical support systems for time-pressured clinicians.           

Biography: 

Anton van der Vegt is currently a full-time PhD student in the School of ITEE, The University of Queensland, jointly supervised with CSIRO/AEHRC, and he has previously completed a B.Sc. in Computer Science and B.E. in Mechanical Engineering at Sydney University.