Learning Across Boundaries: Statistical and Machine Learning Methods for Biomedical Data Fusion

1258 

Invited Paper Session 

Lu Xia (1), Guanqun Cao (1), Haolei Weng (1), Lu Tang (2), Ke Zhu (3), Lu Xia (1)

(1) Michigan State University, N/A, (2) University of Pittsburgh, N/A, (3) NCSU and Duke, N/A

Haolei Weng  
Michigan State University

Guanqun Cao  
Michigan State University

Lu Xia  
Michigan State University

Lu Tang  
University of Pittsburgh

Ke Zhu  
NCSU and Duke

Lu Xia  
Michigan State University

Statistical science is increasingly called upon to integrate knowledge across diverse sources of data, particularly in biomedical research where evidence rarely comes from a single pristine source. Randomized controlled trials, observational cohorts, and large-scale real-world data all provide critical information, but integrating these diverse sources while safeguarding against heterogeneity and negative transfer requires careful methodological innovation. This invited session, "Learning Across Boundaries: Statistical and Machine Learning Methods for Biomedical Data Fusion", brings together recent advances at the interface of statistics and machine learning that address this pressing challenge head-on.

The three presentations will showcase complementary approaches to data fusion, each addressing fundamental barriers to borrowing strength across heterogeneous datasets.

1. The first presentation, "Transfer Learning for Linear Regression with Mismatched Covariates" by Dr. Lu Tang, proposes a novel transfer learning method to develop optimal linear predictors in linear regression using datasets with differing sets of predictors. The proposed method aims to address the challenges of distributional difference and mismatched covariate spaces simultaneously by leveraging covariance from multimodal data and fusion learning.

2. The second presentation, "Robust Estimation and Inference in Hybrid Controlled Trials" by Dr. Ke Zhu, introduces conformal selective borrowing, a novel framework with automatic tuning and randomization inference that adaptively incorporates external real-world data in randomized controlled trials. This approach integrates conformal prediction techniques from the machine learning community and classical randomization principles to improve power while preserving valid inference.

3. The third presentation, "Active Unsupervised Domain Adaptation for Imaging Analysis" by Dr. Lu Xia, proposes a deep learning approach that incorporates active learning strategies into unsupervised domain adaptation, which transfers knowledge from a labeled source dataset to an unlabeled target dataset despite distributional differences, with applications to imaging analysis. Via selective queries of the most informative target samples, this work provides a principled way to decide whether the source data is transferable to the target problem.

Together, these presentations chart a unified vision for robust, transferable, and safe data integration. The discussant, Dr. Guanqun Cao, will synthesize insights across presentations, emphasizing shared principles and pointing to promising directions for theory and practice.

This session is timely, and these advances resonate broadly with statisticians, data scientists, and applied researchers, making the session highly appealing to a diverse JSM audience. The rise of hybrid controlled trials has placed statistical methodology at the center of regulatory discussions, while transfer learning with mismatched covariates directly addresses the challenges faced by biomedical data consortia. Meanwhile, domain adaptation and negative transfer have emerged as critical concerns as machine learning methods move from simulation to practice.

Aligned with the JSM 2026 theme "Communities in Action: Advancing Society", this session demonstrates how statistical and machine learning communities can work across disciplinary and data boundaries to advance biomedical science and societal well-being.

Biometrics Section ³

ENAR ¹

Section on Statistical Learning and Data Science ²