Interactive Machine Learning with Applications in Health Informatics

Abstract

Recent years have witnessed unprecedented growth of health data, including millions of biomedical research publications, electronic health records, patient discussions on health forums and social media, fitness tracker trajectories, and genome sequences. Information retrieval and machine learning techniques are powerful tools to unlock invaluable knowledge in these data, yet they need to be guided by human experts. Unlike training machine learning models in other domains, labeling and analyzing health data requires highly specialized expertise, and the time of medical experts is extremely limited. How can we mine big health data with little expert effort? In this dissertation, I develop state-of-the-art interactive machine learning algorithms that bring together human intelligence and machine intelligence in health data mining tasks. By making efficient use of human expert's domain knowledge, we can achieve high-quality solutions with minimal manual effort.

I first introduce a high-recall information retrieval framework that helps human users efficiently harvest not just one but as many relevant documents as possible from a searchable corpus. This is a common need in professional search scenarios such as medical search and literature review. Then I develop two interactive machine learning algorithms that leverage human expert's domain knowledge to combat the curse of "cold start" in active learning, with applications in clinical natural language processing. A consistent empirical observation is that the overall learning process can be reliably accelerated by a knowledge-driven "warm start", followed by machine-initiated active learning. As a theoretical contribution, I propose a general framework for interactive machine learning. Under this framework, a unified optimization objective explains many existing algorithms used in practice, and inspires the design of new algorithms.