Sleep provides important information about brain health, including potential cognitive impairment (CI). Detecting CI early enables timely intervention, but population-scale screening for it remains impractical. Overnight polysomnography (PSG), collected routinely in sleep clinics, offers a non-invasive window into brain health. The George B. Moody PhysioNet Challenge 2026 therefore focused on predicting future CI diagnoses from PSG recordings collected at five U.S. institutions. In response, our team, PhysioWinn, developed a lightweight convolutional network of 79,457 parameters that uses electroencephalogram (EEG) data from six channels and achieved an age-conditioned AUROC of 0.748 on the hidden validation set, ranking 9th of the 100 participating teams (tied with two others). Our interpretation method indicated that the frontal EEG channels (F3 and F4) contributed most to the model's predictions, and that N3 sleep windows were the most informative. This N3 finding is consistent with sleep literature reporting links between slow wave EEG activity and CI. Our study therefore suggests that frontal-region and N3-aware modeling may further improve prediction and generalization.