A regularized semiparametric cure-rate model with high-dimensional imaging data

Structural magnetic resonance imaging (MRI) is informative for studying progression from mild cognitive impairment (MCI) to Alzheimer's disease (AD), but incorporating high-dimensional images into survival models requires methods that accommodate irregular brain domains and yield interpretable spatial effects. We propose a regularized semiparametric mixture cure-rate model for high-dimensional imaging predictors that separates susceptibility to AD conversion from time to AD among susceptible subjects. Baseline MRI images are represented by bivariate Bernstein spline basis functions over a triangulation of the brain domain, and sparsity-inducing penalties are imposed on imaging effects in both incidence and latency components. This construction enables direct estimation of localized coefficient functions describing how image regions are associated with susceptibility and conversion time. We develop an expectation-maximization algorithm that updates latent susceptibility probabilities and solves penalized logistic and Cox regression subproblems, together with an efficient tuning strategy. Simulations show that the method can recover localized active regions and provide useful incidence and latency discrimination. An application to Alzheimer's Disease Neuroimaging Initiative data demonstrates susceptibility discrimination and interpretable brain-region coefficient maps.