Many methods have been proposed to control the false discovery rate (FDR) in feature selection, yet challenges remain when the relationship between predictors and the outcome is complex. We propose a prediction-based data-splitting procedure that can leverage flexible predictive models while controlling the FDR. The data are divided into three parts: one for model training and two for constructing the mirror statistic. For each feature, we construct a negative control using the remaining features without incorporating outcome information and define a test statistic based on the resulting change in predictive performance. Under the null, such a perturbation is approximately equally likely to increase or decrease predictive performance, yielding a statistic symmetric about zero. In contrast, perturbing an informative feature tends to reduce predictive performance, resulting in a large positive statistic. Features are selected by thresholding these statistics, with false discoveries estimated from the corresponding negative tail to control the FDR. We apply the proposed method to liver function microbiome data, where high-dimensional, correlated features and potentially complex feature--outcome relationships make reliable type I error control particularly important for reproducible scientific discovery. Through simulation studies and the real-data application, we demonstrate that the proposed framework provides robust FDR control and can be integrated with state-of-the-art machine-learning models to accommodate complex relationships between features and the outcome.

翻译：暂无翻译