The George B. Moody PhysioNet Challenge 2026 focuses on early warning of cognitive impairment by developing algorithms to predict the risk of future cognitive disorders from polysomnography (PSG). Our team, Med_YNNU, proposes a novel end-to-end deep learning framework for cognitive impairment risk prediction, integrating a multi-scale dilated residual convolutional neural network (MS-DRCNN) and a hierarchical spatio-temporal attention network (H-STAN) with adaptive feature calibration. First, we implement a multi-stage preprocessing pipeline, including wavelet threshold denoising, adaptive artifact removal, and z-score normalization, to eliminate noise, motion artifacts, and baseline drift in raw PSG signals, thereby enhancing the representation of sleep physiological activity. Second, we employ the MS-DRCNN as the feature extraction backbone: it stacks multi-scale dilated residual convolutional blocks to capture hierarchical spatiotemporal features from multi-modal sleep signals. Third, we design an adaptive dynamic weight learning module to dynamically mine the non-linear correlation between multi-modal physiological signals and sleep stages, alleviating distribution shift caused by individual differences in sleep architecture. Finally, we introduce the H-STAN to jointly model the long-range temporal dependencies across time steps and cross-channel spatial correlations of the extracted features, adaptively assigning learnable weights to critical spatiotemporal features to suppress redundant information and further improve the overall feature representation capability. In addition, we adopt a weighted focal loss function with class-balanced regularization to alleviate the severe class imbalance of the dataset. On the official training dataset of the challenge, the proposed framework yields an average score of 0.401 based on the official challenge metric through five-fold cross-validation. Our team, Med_YNNU, achieves a challenge score of 0.448 on the held-out validation dataset, ranking 81th among over 92 participating teams.