Feature selection aims to explore the characteristics of a problem that is under investigation instead of focusing on extracting (deep) features or classification tasks. The pending issues being explored are as follows: 1) to minimize the interference of uncertain irrelevant features and 2) to construct the (full) set of relevant features as an individual feature that might be too weak, such as those extracted from bio-signals, for example, electroencephalograms (EEGs). This study fosters an adaptive feature selection approach with the Shapley value and hypothetical testing (abbrev. ShapHT+) via adaptive relevance evaluation. The tree SHAP (SHapley Additive exPlanations) method first can be used to quantify the importance of each candidate feature. An adaptive threshold is then derived to evaluate the feature’s relevance to a priori information, and all relevant features are then selected through the hypothesis testing for binomial distribution. The benchmarks indicate that ShapHT+ significantly outperforms its mainstream counterparts in ruling out interferences and the efficiency and accuracy in selecting all relevant features. The case study of EEG feature engineering for autism spectrum disorder (ASD) evaluation indicates that the features selected by ShapHT+ 1) can achieve the highest classification accuracy (82.44%) and 2) well match the observations in recent ASD research.