Polysomnography (PSG) provides a rich, multichannel record of physiological dynamics during sleep, offering untapped potential for early detection of cognitive impairment. Emerging evidence links subtle alterations in sleep oscillations to preclinical neurodegeneration. To leverage this information, we developed a graph neural network (GNN)-based framework that models PSG data as a structured graph, where each physiological channel constitutes a node and functional relationships are encoded as edges. The signals were resampled to 50 Hz and segmented into twelve 30-minute patches to standardize the 6-hour representation, with shorter recordings zero-padded and longer ones cropped. A 1D-convolutional neural network extracted signal embeddings from each patch, which were enriched with algorithmic annotations and demographic features. These multimodal node features were processed by three graph convolutional layers, followed by a fully connected layer producing a binary prediction of cognitive impairment. The model was trained on the challenge's small dataset. Five-fold cross-validation guided hyperparameter selection, with the final model submitted using a batch size of 4 due to computational constraints. On the hidden validation set, the model achieved AUROC of 0.538, indicating weak but non-random discrimination. While accuracy reached 0.861, the F-measure of 0.092 confirms poor sensitivity and precision, underscoring the difficulty of reliable prediction under current conditions. Although the current model demonstrates limited predictive power, it establishes a robust, generalizable framework for modeling heterogeneous, multi-institutional PSG data using GNNs. Future work will focus on refining feature integration, exploring alternative graph structures, and leveraging explainability tools such as GNNExplainer to identify biologically plausible patterns and guide model pruning, potentially improving performance, interpretability, and clinical relevance.