Improving trust in complex genomic data analytics with explainable artificial intelligence

Abstract

The main goal of this thesis is to improve trust in genomic data analytics using designed methods to identify the current explainable AI (XAI) gaps, get domain users’ input and develop novel XAI models. To achieve this goal, we ran a pilot qualitative study with domain users to find domain users' needs and expectations. Then, a narrative review was presented to identify innovation trends and scientific gaps in immersive analysis. The user-centred design was applied to develop an immersive visual analytics system for an observing oncology data prototype. A usability study was conducted with domain users in this prototype to evaluate and identify the interpretability needs. The usability study helped to improve the prototype, including the XAI model development. Based on the results, two novel explainable AI models are developed to explain non-linear project results and rule extraction for better accuracy. The thesis has three contributions to the knowledge: i) insights of immersive analytics and domain expectations, including an overview picture of the immersive analytics and the domain expectations for XAI in genomic data analysis, ii) a method for explaining rule-based supervised ML results and iii) a method for explaining and visualising unsupervised ML results. The research potentially improves the users’ trust in predictive analytics and opens a new and exciting path to aid discovery for disease diagnostics and precision treatments. Technically, visual analytics are used to illustrate complex cancer data, explain AI models, and show interactive predictive results in meaningful and dynamic ways to guide the effective treatment decisions in the cohort.

Date of Award	2024
Original language	English
Awarding Institution	Western Sydney University
Supervisor	Quang Vinh Nguyen (Supervisor), Simeon Simoff (Supervisor), Paul J. Kennedy (Supervisor) & Daniel Catchpoole (Supervisor)