Sleep signals provide a non-invasive window into neurological health and may enable early identification of individuals at risk of cognitive impairment. Team Everybody Hearts (REM) presents a baseline framework for predicting future cognitive impairment from polysomnography (PSG) recordings. Individual signals are resampled to a uniform frequency, normalised, and segmented into fixed-length 5-minute chunks. Each multimodal chunk is encoded into a latent representation using the SleepFM foundation model, and chunk embeddings are aggregated across the night with a masked bi-LSTM to accommodate variable-length recordings. The resulting PSG representation is combined with learnt demographic embeddings for final prediction. Our current baseline is trained only on labelled data. Using cross-validation on the training set, our method achieves an AUROC of 0.817, AUPRC of 0.777, accuracy of 0.798, and F-measure of 0.820. However, performance on the challenge validation set is substantially lower (AUROC 0.471, AUPRC 0.035, accuracy 0.107, F-measure 0.077). This baseline highlights several key challenges, including limited labelled data, severe class imbalance (positive prevalence 5–15%), and domain shifts across acquisition sites. These results motivate several next steps for the official phase, including fuller use of algorithmic sleep annotations, moving from a plain binary classification objective toward time-to-event learning based on the event-time information available in the challenge data, and improved use of unlabelled PSG through virtual adversarial training and pseudo-labelling to enhance robustness under cross-site distribution shift.