This project will develop new efficient algorithms for large-scale nonparametric Bayesian inference in the context of medical prediction problems. The computational techniques will be applied to discover intermediary pathways, known as pathophenotypes, that can be used to predict cardiovascular disease expression and events based on electronic health record and genomic data. While many articles have discussed the promise of electronic health records to improve drug surveillance and efficacy, the advanced statistical methodologies needed to realize this promise requires innovation in efficient computational techniques specially designed for large datasets and modern computer architectures including distributed computing and dealing with streaming data. The contributions of the proposed work will result both in contributions to computer science and statistics as well as in direct impacts for safer, higher-quality healthcare. The work on scalable Bayesian inference will be of general interest to wide breadth of statistics, operations research, economics, and machine learning communities working on efficient inference techniques.
Each of the technical research objectives is inspired by problems in analyzing large medical data, and thus the proposed work will have direct implications for improved healthcare through more timely predictions of adverse drug events and patient outcomes (well-aligned with NSF?s mission to improve national health, prosperity and welfare). The fellowship will allow the researcher to build on her established teaching record through co-instructing a machine learning course at Harvard, supervising 1-2 masters? students on projects related to the proposed research, and high school outreach through the MIT Educational Studies Program.
