Leveraging large-scale pretrained sleep models via multi-modal fusion for early prediction of cognitive impairment

Alvise Dei Rossi1, Giuliana Monachino2, Alessia Calzoni3, Beatrice Zanchi4, Luigi Fiorillo2
1SUPSI, USI, 2SUPSI, 3University of Brescia, 4University of Applied Sciences and Arts of Southern Switzerland


Abstract

As part of the George B. Moody PhysioNet Challenge 2026, our team, MeDiTechDreamers, developed a framework for predicting a future diagnosis of cognitive impairment from polysomnography (PSG) using representations from neural sleep-staging models pretrained on large, heterogeneous PSG datasets. Modality-specific single-channel models independently extracted hypnodensity sequences from available electroencephalographic, electrooculographic, chin electromyographic, and electrocardiographic signals. The sequences were processed by a bidirectional long short-term memory network with attention pooling to produce a record-level embedding, which was fused with demographic, algorithmically derived, and engineered physiological features. Under leave-one-cohort-out cross-validation on the public training data, the selected configuration achieved a mean Challenge score of 0.650. On the hidden validation set, our method achieved a Challenge score of 0.702 (ranked 28th of 97 teams) and a prevalence-based reward of 0.242 (ranked 7th of 97 teams). These results suggest that modality-specific hypnodensity sequences integrated with complementary static features provide a practical basis for predicting future cognitive impairment from a single baseline PSG.