RIP: Optimal Clinical Trials for Personalizing Medical Care: The Expected Value of Oversampling Information

Title: Optimal Clinical Trials for Personalizing Medical Care: The Expected Value of Oversampling Information

Abstract: Personalizing patient care frequently involves selecting treatments based on risk predictions, with patients at lower risk being recommended less aggressive treatment or no treatment at all. Because risk predictions are uncertain, treatments selected based on them may not be optimal. Collecting additional information could, therefore, be valuable. Designing studies to collect the needed information most efficiently is important given that studies have become increasingly expensive to conduct. Methods exist to support such endeavors (i.e., expected value of sample information (EVSI)). However, to date, EVSI calculations consider only studies that estimate overall population means efficiently. Studies to support the personalization of medical care require deciding how much information to collect about which patient subgroups: on which locations along the risk spectrum should studies focus? We develop the Expected Value of Oversampling Information (EVOSI) framework by introducing EVSI into a previously proposed Expected Value of Individualized Care (EVIC) framework, narrowing prediction uncertainty where doing so maximally increases the overall value of individualized treatment choices. With the EVOSI framework, we analyze the features of patient risk subgroups that increase the value of oversampling and conduct numerical simulations to consider trade-offs between these features (i.e., how much to oversample). Results show that studies designed with EVOSI can be expected to achieve more value than those designed with EVSI at a given sample size or the same expected value as EVSI at a smaller sample size. Features that determine optimal oversampling in EVOSI include: subgroup prevalence; one’s priors on the relative distance to the threshold risk at which one should initiate treatment for each subgroup; the amount of uncertainty in these priors for each subgroup; and the amount of available sample for the new study (i.e., the new study’s total sample budget). Deciding to personalize medical care based on patients’ predicted but uncertain risks also requires decisions about collecting additional information. Such risk heterogeneity calls for an EVOSI analysis if, given available sample, the value of personalized care is to be maximized.

Please note: All research in progress seminars are off-the-record unless otherwise noted. Any information about methodology and/or results are embargoed until publication.