Background: Polysomnography captures four biological signatures relevant to preclinical neurodegeneration: impaired glymphatic clearance (fragmented N3 delta activity), thalamocortical decoupling (reduced spindle density and slow-oscillation coupling), locus coeruleus dysfunction (micro-arousals with autonomic instability), and hippocampal hypoxia (REM/N3 respiratory events). These failure modes are sparse across a full night, motivating attention-based architectures that localize pathological epochs.
Methods: A site-invariant framework was developed for patient-level PSG prediction. The primary model was a stacked gradient-boosted ensemble (LightGBM, XGBoost, ExtraTrees) trained on 234 whole-night features capturing sleep architecture, transition dynamics, hypoxic burden, and Complete AI Sleep Report (CAISR)-derived event indices, using site-stratified GroupKFold with held-out sites to mitigate leakage. A distinct gated-attention Multi-Instance Learning (MIL) architecture was evaluated to model sparse and temporally localized signatures of neurodegeneration (e.g., spindle–slow oscillation decoupling, N3 fragmentation, autonomic instability), representing each PSG as a bag of 30-s epochs encoding relative EEG spectral power, spindle density/coupling, and interval-based autonomic dynamics. A hybrid variant combining both architectures via AUROC-optimized late fusion was assessed to characterize epoch-level and whole-night complementarity. All inputs were restricted to site-invariant representations, with demographic covariates excluded to reduce confounding.
Results: The stacked ensemble achieved the best overall performance among evaluated models (Challenge score 0.649; 5-fold cross-validation AUROC 0.683). In contrast, the MIL model exhibited higher cross-site variability (AUROC 0.55–0.65), indicating instability in learned epoch-level attention under limited sample size (n=622).
Conclusion: Site-invariant, whole-night features provide robust PSG-based cognitive risk prediction under cross-site evaluation. While biologically-guided MIL localizes pathological epochs, its generalization is limited at current sample sizes, with performance driven by global summaries. These findings emphasize site-invariant design and the need for stronger pretrained representations for effective epoch-level modeling. Future work will incorporate Short-Time Fourier Transform (STFT)-based Conv2D embeddings and pretraining on large-scale unlabeled sleep data to decouple representation learning from classification and stabilize attention.