The 2026 George B. Moody's PhysioNet Challenge aims to predict future cognitive impairments based on overnight polysomnography (PSG). Our team, Med_YNNU, developed a Dense Multi-Scale Time-Frequency Fusion and Hierarchical Attention Network (DFHANet) to analyze five PSG modalities in the time and frequency domains. We filter, resample, and normalize five typical signals, then segment the signals and extract power spectral features from each segment. We employ five independent temporal-domain dense multi-scale CNNs and five independent frequency-domain CNNs as modality-specific encoders to extract segment-level temporal and frequency domain features from these five signals. The multimodal embeddings are aggregated sequentially through a branch-level multi-head self-attention mechanism, a two-branch 1:1 connection, and two final multi-head self-attention layers, followed by efficient segmentation mask mean pooling, and finally connected to a fully connected classification head. The model is trained and evaluated using subject-grouped five-fold cross-validation. Ultimately, our team Med_YNNU achieved a cross-validation score of 0.664 on the public training set and an official score of 0.646 on the official hidden validation set, ranking 49th out of 102 participating teams.