A1955
Title: Signal recoverability: A unified measure for proxy modality validation
Authors: Suin Kim - Seoul National University (Korea, South) [presenting]
Yoonsuh Jung - Korea University (Korea, South)
Abstract: A signal recoverability functional is formalized to quantify how much of the predictive signal in an expensive primary modality for an outcome is recoverable from a cheaper secondary modality, defined as the fraction of the variance of the primary-modality conditional mean of the outcome that is captured by the secondary modality. Three estimators are constructed from Neyman-orthogonal scores with cross-fitting, differing in their structural assumptions on the numerator: a fully nonparametric two-stage debiased estimator, a quadratic-form estimator imposing a finite-dimensional basis on the secondary modality, and a cross-covariance estimator with matched bases on both modalities. For each estimator, orthogonality is established, exact second-order bias identities are derived, and nuisance-rate conditions for asymptotic normality are provided. The framework is illustrated on recovering the cerebrospinal fluid, amyloid PET, and tau PET biomarker signals for cognitive impairment from an inexpensive plasma panel, where the functional ranks candidate plasma-marker subsets by recoverability and yields a debiased selection of the panel that best reconstructs each primary biomarker signal.
