Machine learning methodology for sequential decision support from large-scale longitudinal data

点击翻译按钮获取中文摘要

Our proposed research will develop computer science and statistical methods for turning large, rich data sets into human-actionable knowledge by providing sequential decision support. It builds on our previous work, which developed related methods that focused on smaller, highly-curated datasets. Using large data sources presents two major challenges: processing large datasets in a timely manner, and effectively communicating the results to an end-user when data are rich and complicated. Our proposed research will tackle both of these issues in turn. Our strategy is to develop new machine learning models that make decisions based both on inherent properties of the items in the data set ("features") and on how items in the dataset are similar or different (through "kernels".) This approach will lead to efficient data processing through convex optimization, and will lead to interpretable results presented through visualization techniques. These methods combined will be able to use large, previously un-tapped data sets to provide sequential decision support.Making good decisions in the present often requires knowledge of potential decisions to be made in the future -- we say a decision is "non-myopic" if it is made based on knowledge of the potential for future decision-making. Non-myopic decision making is important in many application areas; for example, are now seeing the development of electronic data sources that record how thousands or even millions of patients respond to different sequences of treatments over time, and these have the potential to inform evidence-based non-myopic medical decision making more effectively than previous studies. However, rigorous analysis techniques for extracting evidence from this data and effectively presenting it to physicians -- for example through personalized treatment recommendations -- are still in their infancy. Computer science methods in reinforcement learning and machine learning have enormous potential, but in many ways are not suited to providing the evidence required in this setting. Our long-term research goal is to develop reinforcement learning and machine learning techniques so they can be applied to these new sources of sequential medical data, and can in turn provide doctors with the best available evidence for non-myopic decision making. As the depth and breadth of available sequential medical data increases, the methods I will develop will improve the delivery of health care in Canada by providing our medical doctors with new, high-quality evidence to aid them in choosing the best treatments for their patients.