Anticipating, Extracting, and Leveraging Information in Clinical Decision-Making

A principal challenge in machine learning for clinical decision support is the limited availability of high-quality data: Experimental data is hard to acquire and hence scarce, while observational data from the clinic is more abundant but may contain unintended biases. Given that available data is limited either in quantity or quality, it becomes crucial to understand the full extent of its information content - especially in relation to clinical decisions we wish to support. This thesis aims to develop such understanding by studying three modes of handling information in decision-making: (i) anticipating the arrival of future information when making present decisions, (ii) extracting the information encoded in past decisions in an interpretable form, and (iii) leveraging such information to perform other downstream tasks. For anticipating, we consider subpopulation selection in adaptive clinical trials, and formulate this setting as a new type of optimal stopping/switching problem called the optimal commitment problem (OCP). By theoretically analyzing OCP, we discover that, when adapting a trial, decision-makers should factor in the informational value of conducting the trial - beyond just the potential return of its success. For extracting, we tackle the challenge of obtaining transparent representations of an expert's decision-making process purely from demonstrations of their behavior, thereby making the knowledge behind their actions accessible to others. Our efforts establish a new agenda for policy learning focused on understanding - as opposed to merely imitating - human decision-making. We introduce two novel models: interpretable policy learning (Interpole), which explains human actions through decision dynamics and decision boundaries, and lexicographically-ordered reward inference (LORI), which explains human preferences through lexicographically-prioritized objectives. For leveraging, we give two examples of how knowledge extracted from expert demonstrations can inform other downstream tasks: As the first example, we develop inverse contextual bandits (ICB), a method for learning how behavior evolves over time, and show that ICB can help assess the impact of new medical guidelines on actual clinical practice. As the second example, we define the notion of expertise, an information-theoretic measure of how knowledgeable a policy is of its environment, and show that identifying the prominent type of expertise present in a dataset can inform model selection for treatment effect estimation.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2023

Author dc:creator

• Huyuk, Alihan

Advisor dc:contributor.advisor

• van der Schaar, Mihaela

• Clinical Decision-Making
• Machine Learning

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/ed950cf5-0ca3-42f0-a10d-6ccfadbda097/download
• https://www.rioxx.net/licenses/all-rights-reserved/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.112538

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/374493