For this installment in our newly established Biostatistics Seminar Series, Dr. Yuanjia Wang from Columbia University will be giving her presentation on "Domain Adaptation for Learning Individualized Treatment Strategies from Randomized Trials Supplemented by Electronic Health Records".
ABSTRACT: Individualized treatment rules (ITRs) tailor medical treatments according to patient-specific features in order to improve treatment outcomes. Recently, methods for estimating ITRs from randomized controlled trials (RCTs) and observational studies (e.g. electronic health records, EHRs) have increasingly received attention. RCTs are often conducted under stringent inclusion/exclusion criterion, limiting the generalizability of ITRs learned from RCT to the broader, real world patient population. Since EHRs document treatment prescriptions in the real world, transferring information learned from EHRs to RCTs, if done appropriately, could improve the performance of ITRs, especially in the general population. The strategy of transferring information is referred as domain adaptation. In this work, we propose a kernel-weighted matched learning method and domain adaptation to enhance learning ITRs from RCT supplemented by information from EHRs. Specifically, we achieve domain adaptation through pre-training super features from the EHRs and projecting on the RCTs to augment its feature space. Under this framework, two machine learning methods including Q-learning and Kernel-Weighted Matched Learning (M-learning) are applied to observational EHRs and randomized trials to accurately estimate ITRs. To maintain the virtue of randomization and protect ITRs from unmeasured confounders, the ITRs mainly rely on RCT data but uses EHRs as auxiliary information. Simulation studies demonstrate superiority our proposed method when covariate distributions differ between two domains of data source. Lastly, in real world application, we apply our framework to transfer information learned from EHRs of type 2 diabetes (T2D) patients to improve individualized insulin therapies learned from a RCT.
ABOUT THE PRESENTER: Dr. Wang is an Associate Professor in the Department of Biostatistics and Department of Psychiatry at Columbia University, and a core member of the Division of Biostatistics at New York State Psychiatry Institute. She was elected as a Fellow of the American Statistical Association (ASA) in 2016. Dr. Wang works on developing data-driven approaches to explore relationship between biomarkers, clinical markers, and health outcomes to assist discoveries in disease etiology and increase diagnostic capabilities of psychiatric and neurological diseases. Her methodological interests include statistical learning, analytics for personalized medicine, network analysis, and novel design and analysis of clinical trials. Her substantive research area of interest includes psychiatric disorders and neurological disorders.
Registration required. Please disregard the closed webform below and instead RSVP here.