Collaborative Research: Integrating Multimodal Data in AI for Precision Medicine

Modern medicine collects many kinds of information about individual patients, including medical images, genetic data, and routine clinical measurements, but these data are rarely combined in ways clinicians can fully trust and understand when making treatment decisions. This project develops transparent and trustworthy artificial intelligence and statistical methods that bring these diverse data together to guide individualized and adaptive medical care. The central challenge is turning multimodal, large, complex, and noisy patient data into clear, reliable guidance that supports better treatment choices and adapts as a patient's condition changes over time. By creating tools whose recommendations can be explained and whose reliability can be measured, the project helps make medical artificial intelligence safer and more credible for high-stakes health decisions. The methods will be demonstrated using real patient data, including data on Alzheimer's disease, to support early detection, risk stratification, and more personalized treatment. The project advances the national health and welfare by strengthening the foundations of dependable medical artificial intelligence, produces freely available software, and trains graduate and undergraduate students in interdisciplinary research. This project develops a unified statistical and machine learning framework for interpretable and trustworthy integration of multimodal biomedical data, including medical imaging, genomics, and clinical variables, into individualized and sequential treatment decision making. The work targets factors that limit the joint use of such data, including high dimensionality, complex dependence, heterogeneity, and noise, and emphasizes valid uncertainty quantification, theoretical rigor, and scalability via three coordinated efforts. The first effort develops interpretable statistical learning for single-stage individualized treatment rules, emphasizing inference and uncertainty quantification. The second effort develops interpretable direct learning, or policy search, that estimates optimal treatment rules without explicit outcome modeling, improving robustness in complex applications. The third effort develops interpretable reinforcement learning for dynamic treatment regimes, enabling sequential policies that adapt to evolving patient status, including longitudinal biomarkers. The framework unifies semiparametric learning, functional and structured data representations, direct policy learning, and reinforcement learning, and establishes theoretical guarantees such as large-sample properties, robustness, and uncertainty quantification. The methods will be evaluated using real-world data, including the Alzheimer's Disease Neuroimaging Initiative, to demonstrate relevance to early detection, risk stratification, and treatment guidance. The project will also deliver scalable algorithms and open-source software supporting reproducible research and broad adoption. 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.