Predicting future cognitive impairment (CI) from polysomnography (PSG) requires a model that generalizes to a recording site it has never seen, at low prevalence. I built an 899-feature gradient-boosted stack combining hand-crafted clinical and spectral features with embeddings from a frozen pretrained sleep transformer (LPSGM) and a spectrogram autoencoder (SGAE) trained on the Challenge data. Submissions were made as team CCA_BH and reached an age-conditioned AUROC (AC-AUROC) of 0.742 and a prevalence-based reward of 0.240 on the official validation set, placing 38th of 514 scored entries (with a team rank of 13 out of 103).