Machine learning-based methods for phenotyping dementia patients from electronic health record data

PROJECT SUMMARY/ABSTRACT Candidate: Dr. Roy Adams applies for this K25 Mentored Quantitative Research Career Development Award with the goal of building a productive independent research career as a methodologist focused on developing electronic health record (EHR)-based models and tools to improve our understanding of Alzheimer’s disease and related dementias (ADRDs). Dr. Adams brings with him excellent training in computational methods for observational health data but lacks expertise in ADRDs and the methods used to study them. “Big data” is powerful but understanding the context surrounding the data is essential for knowing the limits of the data and avoiding bias. The K25 training will support Dr. Adams in becoming an independent ADRD researcher by allowing him to: (1) develop an understanding of dementia biology and care, (2) gain expertise in the methods used to model psychiatric measurements, (3) gain exposure to the study of ADRDs from observational data, and (4) form a network of collaborators in clinical ADRD research. These training aims will be accomplished through in-person clinical exposure, didactic courses, directed readings and journal groups, and participation in professional research networks. Research and Environment: Phenotyping is an essential step of most EHR-based studies of ADRDs. Due to common sources of error – such as fragmented care and selection bias – phenotyping ADRDs in EHR data remains a challenge. Recent advances in machine learning present a potential way to account for these sources of bias in high-dimensional EHR data by combining multiple proxies for the phenotype of interest, while explicitly modeling the error and bias in each proxy. However, these methods remain limited and methodological development is needed before they can be applied to ADRD data without risking substantial bias. The proposed research focuses on developing these methods to extract two types of EHR-based phenotypes of ADRD: a binary phenotype indicating whether a patient has dementia and a continuous phenotype measuring the severity of that dementia. Dr. Adams will apply these methods to a large database of Johns Hopkins EHRs and validate them using a combination of data from a memory center, data from a parallel ongoing longitudinal study of ADRDs, and assessments of patient severity based on chart review. This work will take advantage of a unique combination of resources available through the Johns Hopkins Alzheimer’s Disease Research Center, the Richman Family Precision Medicine Center of Excellence in Alzheimer’s Disease, and the Johns Hopkins inHealth Precision Medicine initiative. Further, this research will provide Dr. Adams with valuable experience working with ADRD patient data, set the foundation for future methodological work, and generate methods that can be directly applied to several planned and ongoing ADRD precision medicine studies at Johns Hopkins.