Sleep disturbances are among the earliest detectable biomarkers of neurodegenerative disease, yet translating overnight polysomnography (PSG) recordings into reliable long-term cognitive risk scores remains an open challenge. In the context of the George B. Moody PhysioNet 2026 Challenge, we research whether large pretrained foundation models can be utilized instead of implementing costly task-specific training pipelines for predicting future cognitive impairment from a single baseline PSG recording.
Our proposed method transfers representations from a large pretrained sleep foundation model, where raw multi-channel PSG signals are segmented into five-minute windows spanning four physiological modalities: brain and ocular activity (EEG/EOG), respiratory effort, cardiac rhythm, and muscular activity. While each window is independently encoded by the backbone into a 128-dimensional embedding, we train an attention pooler that re-weights individual windows and produces a single 512-dimensional recording-level representation. This pooled embedding is concatenated with 83 hand-crafted features derived from demographics, Hjorth signal statistics, and algorithmic sleep annotations (CAISR), yielding a 595-dimensional feature vector that is then passed to a gradient-boosted tree classifier.
In our experiments, attention-pooled embeddings combined with handcrafted features outperformed all other pooling methods, suggesting that cognitively informative PSG features are concentrated in specific temporal epochs rather than uniformly distributed across the night. In the unofficial challenge phase, our team DeepHypnos achieved a score of 0.56, placing 96th across all entries.
These results demonstrate that a pretrained sleep foundation model, combined with a temporal attention mechanism and classical ensemble learning, provides a strong and computationally efficient baseline for PSG-based cognitive impairment prediction.