Collaborative Research: Causal Learning with High-dimensional Imaging Outcomes: Methods, Theory, and Algorithms

About This Grant

The analysis of imaging outcomes is a dynamic and rapidly evolving research field, driven by the growing accessibility of large-scale biomedical imaging databases. Imaging data, often characterized as functional data, presents unique opportunities and challenges for statistical analysis. Existing methods, however, are insufficient for handling the computational demands of large-scale medical imaging data or addressing issues such as unmeasured confounding and population heterogeneity in causal analysis. This research will develop advanced statistical tools to overcome these critical hurdles. By developing new techniques that efficiently process large-scale imaging information and provide more accurate causal insights, this work will advance national interests in scientific innovation and evidence-based decision-making. It will promote scientific progress in a vital area of imaging data analysis and aims to advance public health by enabling a deeper understanding of treatment effects from observational studies. The developed data analytics tools also have broad applicability across various fields, including aging research, digital health, and plant science, addressing challenges faced by modern society. Furthermore, the project will benefit the broader research community through the release of freely available software tools and will support STEM education by involving undergraduate and graduate students in hands-on research and integrating project findings into curriculum development. This project aims to develop a general functional data analysis (FDA) framework for analyzing large-scale imaging data and uncovering causal relationships between treatments/exposures and imaging responses. Specifically, the project will address challenges in large-scale observational imaging studies via three aims. First, it will develop functional regression models for imaging responses based on a distributed learning framework, enabling scalable yet accurate estimation and inference. Second, it will introduce an image-on-scalar instrumental variable regression to mitigate confounding bias in observational studies. Third, it will propose an image-on-scalar doubly robust regression method leveraging functional pseudo-outcomes to address population heterogeneity. The proposed methods will be rigorously evaluated using existing imaging studies and are expected to significantly advance the methodology, theory, and computation of FDA and causal inference. Additionally, by releasing open-source software, the project will empower researchers to harness vast amounts of imaging and functional data from publicly available repositories. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.