Stable Neuroimaging Feature Selection via Multi-Knockoff Simultaneous Inference

This study addresses the critical challenge of stable and accurate prediction of cognitive and disease states in neuroimaging analysis, with a focus on early detection of neurodegenerative diseases such as Alzheimer's Disease (AD). Existing knockoff-based feature selection methods, though promising, exhibit instability and limited power to identify relevant features, restricting their practical application in neuroimaging contexts. To overcome these limitations, we propose Stable Simultaneous FDP Bounds with Multi-Knockoffs (StableSim-MK), a novel framework that integrates multi-knockoff selection with simultaneous selective inference. StableSim-MK enhances stability, significantly increases the number of reliably identified biomarkers, and provides rigorous control over the False Discovery Proportion (FDP). Experimental results demonstrate that StableSim-MK identifies a broader set of AD-associated brain regions compared to existing methods, achieving both higher stability in selection and superior predictive accuracy. Notably, the proposed method offers robust and interpretable feature selection while revealing novel insights into brain regions that are overlooked by conventional approaches, thereby establishing meaningful links to cognitive states.