Introduction. PSG captures neural (EEG), ocular/muscular (EOG/EMG), autonomic (ECG/respiration) signals, and sleep macrostructure altered in cognitive decline like MCI/Alzheimer's. Integrating deep learning–derived representations from both supervised and self-supervised large-scale models with clinical characteristics and demographic variables may capture complementary micro- and macro-level sleep characteristics-dynamics, improving robustness to site-specific heterogeneity. Aim: predict cognitive impairment (CI) 3-7 years post-PSG.
Current method. We first build and scale (RobustScaler on each index) time‑invariant "static" features from demographics (age, BMI, sex, race) and twelve CAISR-derived indices (e.g., AHI, arousal, limb-movements, etc.). Physiological signals (EEG, EOG, CHIN, ECG), are pre-processed via IQR-based clipping, resampling to 128 Hz and robust scaling. For each lead-type, we then exploit U-Sleep models from SLEEPYLAND (pre-trained on over 220'000 hours of PSG recordings on sleep staging task) with modality‑specific weights (EEG/ EOG/CHIN/ECG) to extract epoch‑level supervised-representations from an intermediate layer, yielding per‑epoch embeddings for all available channels. For cognitive impairment classification, we combine (i) sequences of per‑epoch SLEEPYLAND embeddings across modalities (randomly extracting one channel per available modality for each sample), processed with a single-layer LSTM and (ii) the static feature vector, concatenated through a dedicated branch, with fusion performed in the final layers. The model is trained with BCE loss, using stratified train/validation/test splits, early stopping, and metrics computed on a held‑out test set.
Results. We achieved AUROC=0.581, AUPRC=0.048, Accuracy=0.644, F-measure=0.093 on the hidden validation set and AUROC=0.569±0.023, AUPRC=0.585±0.017, Accuracy=0.537±0.019, F-measure=0.5±0.043 on internal 5-fold CV.
Next steps. We plan to (i) expand on model derived representations by including SSL encoders (e.g., sleepFM, Self-DANA), whole-night representations (e.g., SLEEPYLAND-derived hypnodensities), and epoch-level CAISR events at different time granularities, (ii) improve the integration of multimodal data and robustness to missing modalities by exploring principled attention-based fusion with modality masking, random modality dropout, FiLM-conditioning, and Coupled Mamba state-space models for multi-modal CI prediction.