Nonparametric methods to address scientific questions in aging integrating real-world and research databases

The National Institute on Aging awarded Kaiser Foundation Hospitals $830,868 on September 3, 2026, under the Aging Research program (CFDA 93.866) to develop nonparametric methods integrating electronic health record and validated research data for aging-related scientific questions.

Kaiser Foundation Research Institute will develop machine learning methods to improve accuracy and precision of inverse-probability weighted analyses and validation study design for causal inference and prediction model evaluation. The work addresses methodological challenges in leveraging error-prone electronic health record data—particularly unstructured clinical text extracted via text-mining algorithms—alongside validated data from smaller cohorts, such as the Eye Adult Changes in Thought study. Specific deliverables include methods that properly incorporate machine learning-based weights by combining advances in generalized raking with targeted learning approaches, and improved design frameworks for validation studies in causal inference and prediction model evaluation. The research aims to provide reliable scientific and clinical evidence for prognostic tool development in eye diseases including age-related macular degeneration and for understanding relationships between ophthalmic biomarkers and Alzheimer's disease and related dementias risk.

Performance occurs in Seattle, Washington, with a period of performance from award date through June 30, 2031. The award is classified as a Project Grant, the standard assistance type under the NIH Aging Research program.

Abstract

Electronic health records (EHR) data collected during routine clinical care are a cost-effective data resource to inform studies of disease pathology and prognosis. In aging research focusing on eye-brain connections, there is great interest in using EHR data to develop prognostic tools for eye diseases (e.g., age-related macular degeneration [AMD]), and to understand the relationship between ophthalmic biomarkers and risk of alzheimer's disease and related dementias (adrd). However, in large cohorts such as the eye adult changes in thought (eye act) study, target variables are often extracted from the EHR, particularly unstructured text in clinical notes, using error-prone text-mining algorithms. As in many other settings, validated data by chart abstraction in eye act are only available on a subset of the cohort. Machine learning methods that leverage detailed data of varying quality can improve accuracy and robustness of analyses to provide reliable scientific and clinical evidence. The proposed research will address three key methodological issues that remain in using machine learning to incorporate error-prone and validated EHR data into analyses to answer scientific questions: 1) improving accuracy and precision of machine-learned inverse-probability weighted (IPW) analyses; and improving design of validation studies for 2) causal inference and 3) prediction model evaluation. Many analyses that combine error-prone and validated data involve IPW analyses of the validated subset. Weights derived using machine learning methods, while useful for prediction, can result in overly variable weights, leading to reduced accuracy or precision of an estimator. This problem can be exacerbated in small samples. We will develop methods that properly incorporate machine learning-based weights, combining advances in generalized raking with targeted learning (e.g., targeted maximum likelihood estimation) for robust and stable weights. When designing data- validation studies, simple random sampling is often inefficient; stratified samples can increase efficiency by validating key subgroups and by utilizing existing optimal design methods that minimize the variance of regression parameter estimates. We will generalize these optimal design approaches to estimate average treatment effects and other targets of causal inference, addressing an important void in the current optimal stratified-design literature. We will further extend these methods to develop the optimal design for evaluating a prediction model and for simultaneously evaluating a model and performing causal inference. By combining advances in machine learning, targeted learning, and survey statistics, we will provide more robust and precise statistical tools to study the eye-brain connection in aging and to generate reliable empirical evidence on actionable factors. Our work using eye act has the potential to transform preventive care for important aging outcomes like AMD and adrd, while also serving as a statistical framework for improved inference and study design more broadly in biomedical research that leverages observational data with missing or error-prone data.