Sleep polysomnography (PSG) records a rich set of physiological signals that provide information well beyond the diagnosis of sleep disorders. Recent evidence suggests that subtle alterations in these signals may precede the onset of cognitive decline, offering a valuable window for early intervention. Additionally, Graph Neural Networks (GNNs) are increasingly used in biosignal analysis because they can effectively capture non-Euclidean structure in data. In the Computing in Cardiology Challenge, our team, QueensOfDreams_UMG, employed a GNN framework to predict cognitive impairment of the multi‑institutional PSG dataset. All PSG channels were resampled to 50 Hz. To reduce computational load, the first hour, dominated by wakefulness, was discarded. The remaining recording was segmented into twelve 30‑minute patches, yielding a 6‑hour representation. Recordings shorter than 6 hours were padded with zeros, longer recordings were cropped. Each channel constitutes a node in a graph, and physiological knowledge guided the definition of spatial edges. Redundant or missing channels were handled by fallback connections. Node features were extracted using a 1D-convolutional neural network consisting of 3 convolutional layers followed by adaptive average pooling. In addition to raw signal embeddings, algorithmic annotations, demographic variables, and statistical descriptors were concatenated. The resulting node embeddings were processed by 3 graph convolutional layers, followed by a fully connected readout that produced a binary prediction of cognitive impairment. Preliminary results by using I0006 and S0001 as training data and I0002 as validation, yielded an AUROC of 0.512 and an accuracy of 0.472. Training on the entire dataset, the Challenge submission system provided an AUROC of 0.573 and an accuracy of 0.438. To enhance interpretability, we aim to apply GNNExplainer to identify influential channels and edge patterns driving model decisions. This approach could enable retraining after pruning less relevant connections, potentially improving performance, robustness, and yielding clinically plausible importance maps.