Aim: We study cross-site generalization in screening for cognitive impairment from polysomnography (PSG), and characterize the limits of a leave-one-site proxy protocol for model selection across families. This work is part of the George B. Moody PhysioNet Challenge 2026, in which we participate as Team Revenger.
Methods: We compare three model families. The first is an epoch-level convolutional recurrent network (CRNN) over Complete AI Sleep Report (CAISR) features. The second is gradient-boosted trees on a hand-crafted 390-dimensional spectral block. The third uses frozen 1024-dimensional embeddings of a sleep-EEG brain-health model, reduced by principal component analysis (PCA) to 64 dimensions and ranked by XGBoost or logistic regression. Model selection uses a leave-one-site split with single-factor changes.
Results: The spectral family achieved the best result: age-conditioned AUROC 0.645 to 0.749 in per-site cross-validation on the public training set and 0.627 on the hidden validation set (rank 209/514).
Conclusion: Cross-site generalization depended on the representation: spectral features transferred best, CAISR event rates drifted most, and frozen embeddings offered no gain. The proxy was accurate only for the spectral family, so selection requires per-family validation.