Aims : This work investigated whether raw PSG waveform morphology and tabular sleep metrics encode complementary information predictive of cognitive impairment (CI), seeking signal features that discriminate preclinical CI from overnight recordings.
Methods : A stacked ensemble combined a one-dimensional CNN on raw PSG waveforms with an XGBoost classifier trained on tabular features. The CNN processed 30-second windows sampled preferentially from N3 and REM stages across nine channels: high-pass filtered EEG (C3, O1, F3, CHIN at 0.5 Hz), bilateral EOG, SpO2 waveform, CAISR algorithmic sleep stage, and an F3-F4 frontal differential. Tabular features included sleep architecture metrics, oxygen saturations, respiratory and arousal indices, limb movement index, CAISR stage probabilities, and EEG spectral band power (delta, theta, alpha). Stacking weights were optimized by sweep on a held-out validation set. BMI emerged as the dominant XGBoost feature but likely reflects both physiological signal and site-level missingness, motivating further investigation. The training dataset was unexpectedly updated mid-unofficial-phase to exclude patients with pre-existing CI diagnoses; model development during the unofficial phase was based on the initial version, explaining the gap between local holdout and official scores. Results: Best official score was AUROC 0.582 (Team Name CCA_BH). On local held-out validation (122 subjects, 50% CI prevalence), CNN achieved AUROC 0.637 and XGBoost 0.597 independently; stacking improved performance to AUROC 0.663 (α=0.4). The CNN's independent contribution suggests waveform morphology encodes CI-relevant signal beyond standard sleep scoring metrics.
Planned extensions : The CNN backbone will be replaced with a ResNet incorporating squeeze-excitation attention, following top-performing PhysioNet teams. A two-stage model aggregating per-window embeddings across the full night using temporal attention will address the limitation of simple window averaging. Pretraining on sleep stage classification using public PSG datasets before CI fine-tuning is planned. Spectral artifact rejection and EKG-based EMG artifact removal are under development.