Sleep is a fundamental physiological process essential for cognitive function, neural recovery, and overall systemic health. Emerging evidence suggests that disturbed sleep patterns are associated with and may precede the onset of cognitive impairment, including Alzheimer's disease or dementia. As part of the 2026 George B. Moody PhysioNet Challenge, we (Team TemporalSync) propose a spectrogram based deep learning framework for predicting future cognitive impairment from raw polysomnography (PSG) data. The model employs a convolutional neural network (CNN) operating on multi-channel mel-spectrograms derived from resampled physiological signals. Each PSG recording is segmented into 16-minute windows and transformed into time–frequency representations, which are combined with patient demographic information to incorporate clinical context. To improve class separability, we adopt a Supervised Contrastive (SupCon) loss during training, encouraging the model to learn compact intra-class representations while maximizing inter-class separation. This objective aims to capture subtle temporal and spectral patterns associated with early cognitive decline. During inference, predictions are generated for individual segments across the full recording and final patient classification is obtained via majority voting. Our approach achieved an AUROC Challenge score of 0.379 during the unofficial phase, ranking 234th out of 244 team runs. This result highlights the difficulty of extracting discriminative features from PSG-based representations and suggests that the proposed method does not yet adequately capture the complex temporal dynamics and informative spectral characteristics of the data. These findings emphasize the need for improved modeling strategies that better exploit long-range temporal dependencies and cross-modal interactions within PSG signals. In particular, future work should move beyond a single CNN architecture toward modality specific encoding with late fusion, enabling each type of physiological signal to be learned independently before integrating complementary information. Advancing such approaches may be critical for enabling more accurate early detection of individuals at risk of cognitive impairment.