Abstract: It is very common to study the statistical significance between mild cognitive impairment (MCI) and normal control (NC) groups with resting-state functional magnetic resonance imaging (rs-fMRI) functional connectivity (FC) and machine learning. However, the significant FC regions from different data or different methods are not entirely consistent. In this paper, we propose a neurological FC-disconnection feature-selection method based on a seed point analysis (SBA) technique. The proposed method studied the FC in a negative and a positive network under the resting state, and tried to find more consistency from the networks than from regions. The experiment used public rs-fMRI data. The results show that the classification accuracy of the proposed feature-selection method is about 8 percentage points higher than the traditional method. After the selected FC features counted in the positive and negative network, the features of the positive network of MCI are about 12% more than those of the negative network. The results prove that there may be an imbalance between the two networks and the lack of normal competition in the psychological activities of the services, leading to the decline of patients' attention.