Tuesday, Aug 4: 10:30 AM - 12:20 PM
6012
Contributed Papers
Thomas M. Menino Convention & Exhibition Center
Room: CC-257A
Main Sponsor
Biometrics Section
Presentations
Multi-institutional clinical registries provide valuable opportunities for disease phenotyping but are often constrained by data privacy requirements that preclude centralized analysis. We propose SOLO-fedLCA, a single-round federated latent class analysis framework that enables multi-center phenotyping without sharing individual-level data. SOLO-fedLCA uses a one-time broadcast of aggregated derivative information to construct an estimator that is asymptotically equivalent to pooled-data analysis, while substantially reducing communication overhead. We apply SOLO-fedLCA to a seven-center peripheral artery disease registry and identify three clinically interpretable phenotypes that differ in systemic atherosclerotic burden and limb ischemic severity. The resulting phenotypes stratify 12-month risk of major adverse limb events. Substantial between-center variation in phenotype prevalence underscores the need for federated methods to support reproducible multi-institutional phenotyping. Overall, SOLO-fedLCA provides a practical and statistically efficient framework for collaborative clinical research under real-world data-sharing constraints.
Keywords
Federated inference
Latent class analysis
One-shot estimation
Majorization–minimization
We study estimation and inference for a finite-dimensional parameter by integrating multiple heterogeneous data sources. Prior data fusion works typically assume that source distributions fully or weakly align with the target distribution via shared conditional distributions that are exchangeable or differ up to an unknown finite-dimensional parameter. However, in many applications source data heterogeneity exceeds these alignment regimes. In this work, we consider efficient estimation under data fusion when some, possibly infinite-dimensional, functionals are exchangeable between the target and source distributions. In this setting, we derive the semiparametric efficiency bound and characterize the efficiency gain by integrating heterogeneous source data. We also provide a general construction of efficient estimators that uses all available data and leverages flexible machine-learning methods. We illustrate our theoretical results with the example of estimating average treatment effect with external controls. We also demonstrate the performance of the proposed estimator via simulation studies.
Keywords
Data fusion
semiparametric efficiency
debiased machine learning
causal inference
Multiple comparative studies with binary outcomes are frequently used in biomedical research to estimate the overall risk difference (ORD) for evaluating drug efficacy and profile safety. However, estimating the ORD is challenging, particularly when trial sizes are small or when studies report zero events. While the ORD is traditionally estimated using a weighted average of summary statistics from 2x2 tables, this approach is often biased because it excludes zero-event studies or relies on continuity corrections. In contrast, model-based methods can more effectively incorporate zero-event studies. In this article, we develop several efficient model-based estimation procedures for the ORD and compare them against weighted average-based methods through simulation studies. Finally, we demonstrate these procedures by analyzing a real-world dataset related to COVID-19 and associated diseases.
Keywords
Binary outcomes
Comparative studies
Estimation procedures
Overall risk difference,
In regression meta-analysis, we often encounter situations where multiple studies use linear regression models, and only a subset of covariates is observed in each study, with different studies using different subsets. In such situations, important confounders are omitted in the studies, and this makes the regression coefficients unidentifiable. Prior methods propose ways to estimate the regression coefficients despite the omission of confounders, but they overlook inherent heterogeneity across studies. We propose a random effects meta-analytic model, accounting for study heterogeneity by viewing the study-specific regression coefficients as iid draws from a normal distribution. The model requires that for each study, we have estimates of the joint distribution of the covariates for that study. These are available from an external source of unlabelled data. We develop consistent and asymptotically normal estimates of the mean and variance of the normal distribution, and obtain closed-form expressions for estimates of the study-specific regression coefficients. We develop an iterative algorithm for estimation and illustrate its performance through simulations and real data analysis.
Keywords
Causal effect
Data integration
Meta-analysis
Observational studies
Random effects
Survival analysis in biomedical research is constrained by data scarcity, inter-institutional heterogeneity, and privacy regulations that prohibit centralized data sharing. Federated learning preserves data locality but typically focuses on learning a single global model and does not directly address limited target-site sample sizes. Transfer learning can mitigate this limitation, yet many existing methods require access to source data or iterative communication. We propose Federated Transfer Learning via Random Survival Forests (FTRSF), a privacy-preserving and communication-efficient framework for decentralized time-to-event analysis. FTRSF integrates federated and transfer learning by training models locally at auxiliary sites and transferring knowledge asymmetrically to a designated target site through pseudo-features derived from auxiliary risk scores. This one-shot approach accommodates censoring and complex risk structures without sharing individual-level data. Simulation studies and a real-data application show that FTRSF improves target-site prediction accuracy, particularly for small or heavily censored cohorts, while remaining robust to moderate inter-site heterogeneity.
Keywords
Confidential data
Decentralized data
Distributed learning
Model sharing
Random survival forests
Causal inference across multiple data sources has the potential to improve the generalizability of scientific findings. Integrating data sources with partially observed outcomes or covariates provides a promising framework for enhancing estimation efficiency and external validity. Under random sampling from a common population, our previous work showed that integrating large incomplete datasets with summary-level data yields efficient, unbiased estimates. In this study, we propose a novel statistical framework for integrating summary-level data with the information of heterogeneous data sources. The proposed method estimates study-specific sampling weights based on the auxiliary information and uses them to recalibrate the estimating equations for the full model parameters. The performance of the proposed method will be evaluated under various sampling designs using simulation studies and applied to the reanalysis of data from U.S. cancer registries and summary-level odds ratio estimates of selected colorectal cancer (CRC) risk factors while relaxing the random sampling assumption.
Keywords
Casual Inference
Data integration
Data fusion
Aggregated data
Propensity score
Sampling weight calibration
Coronary artery bypass grafting (CABG) is performed in nearly 400,000 U.S. patients annually, improving survival and quality of life. However, 30-40% experience complications and 10–15% are re-hospitalized, many preventable with guideline-based care. Current value-based programs are limited by their focus on independent outcomes rather than joint evaluation of related endpoints, inadequate adjustment for patient and social risk factors, and the inability to integrate distributed patient-level data across hospitals. Motivated by these challenges, we develop a federated learning framework that leverages multi-site EHR data to jointly analyze multiple outcomes with robust case-mix adjustment. Our framework first fits a multivariate generalized linear mixed-effects model in a federated manner and then uses resulting estimates to compute directly standardized event rates to assess hospital performance. Validation with simulation studies and real-world CABG cohorts demonstrates the feasibility and utility of our approach while protecting patients' confidentiality, enabling multidimensional hospital performance evaluation using real-world data in distributed research networks.
Keywords
Federated Learning
Multivariate Methods
Hospital Profiling
Electronic Health Record
Mixed-effects Modeling