Mild cognitive impairment (MCI) is a key stage in the diagnosis of senile dementia, and the characteristics of electrical brain (EEG) signals can reflect the cognitive status of MCI patients and help achieve early diagnosis. In the process of EEG feature extraction, most existing studies use fixed time windows to complete segmentation processing for each rhythm of EEG, ignoring the feature differences of different rhythms, thus affecting the diagnostic effect. In view of this problem, a new combined sliding window optimization algorithm is proposed, which improves the construction method of zero model by iterative amplitude adjustment Fourier transform (IAAFT), so as to evaluate the brain dynamic characteristics KPLI. By adopting a variety of sliding window combinations for EEG frequency band signals and guiding them with KPLI indicators, the best sliding window combinations suitable for different frequency bands are obtained. Based on the best sliding window combination, the phase lag index (PLI) is extracted from each band combination, the continuous wavelet transform (CWT) feature is performed, and the MCI diagnosis is realized through the ResNet-MLP dual-channel classification network. The results show that diagnostic classification was achieved for 88 subjects (32 MCI patients, 36 Alzheimer’s disease patients, and 20 healthy control) using a personalized combination band sliding window, and the classification accuracy was 82.2%, which is 10% higher than the classification of fixed window (72.2% classification accuracy is obtained). The results showed that based on the individualized EEG rhythm feature combination, the features of MCI could be better extracted, and the accuracy and specificity of the diagnosis of mild cognitive impairment could be improved, which was an effective EEG feature extraction method.