Data-Driven Human Decision Augmentation with Machine Learning

The increasing availability of observational data has fuelled efforts to enable more personalized decision-making in medicine and healthcare. However, the large volumes and high dimensionality of these datasets present significant challenges for human analysis and evaluation. To address these challenges, the development of advanced decision augmentation systems is imperative. These systems empower humans to make informed decisions by integrating expert knowledge, optimal decision strategies, and predictive evidence of potential outcomes derived from observational datasets. To shed new light on the study of data-driven decision augmentation, this dissertation identifies and investigates four fundamental problems in this area: 1) extraction of expert knowledge from observational data, 2) search for timely and effective decision and sensing strategies, 3) discovery of evidence and insights to inform personalized decision-making, and 4) assessment of cross-domain validity of predictive models and decision strategies. Leveraging theories and techniques in quantitative epistemology, reinforcement learning, predictive clustering, and automated machine learning, we introduce new mathematical formulations, develop novel machine learning models and algorithms, and provide experimental evaluations to demonstrate the practical utility of our proposed solutions to these problems. Specifically, for expert knowledge extraction, we formalize the desiderata for understanding clinical decision-making using a case study on organ transplantation and propose a data-driven framework to identify key risk factors affecting decisions on organ offers in an individualized manner. For the search of effective sensing strategies, we introduce a novel risk-averse formulation of the active sensing task to tackle the continuous-time decision-making problem for longitudinal patient follow-ups. To facilitate evidence-based decision support using longitudinal observations, we develop a predictive clustering algorithm to discover patient phenotypes that associate unique temporal patterns in patient covariates with typical outcomes using frequency domain representations. Finally, to highlight the significance of cross-domain validity of machine learning models in decision augmentation, we provide a detailed analysis with registry data from cystic fibrosis patients of different demographics. By exploring these four fundamental perspectives, this dissertation effectively advances the frontier of data-driven decision augmentation and offers valuable insights for future research in this field.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2024

Author dc:creator

• Qin, Yuchao

Advisor dc:contributor.advisor

• van der Schaar, Mihaela

• machine learning
• decision augmentation

Licence

• https://www.repository.cam.ac.uk/bitstreams/7aad41a7-2430-4af5-8a9e-d6660f24f6b0/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.113278

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/375753