Aims: Sleep is a fundamental biological process closely linked to cognitive function, yet its relationship with cognitive decline remains incompletely understood. This study aims to investigate multimodal polysomnography (PSG) using fractal-based feature extraction to characterize signal complexity, self-similarity, and latent sleep dynamics for predicting mild cognitive impairment (MCI).
Methods: Data from the Human Sleep Project, including demographic information, multimodal PSG signals, sleep staging annotations (30-s epochs), respiratory events, and movement events from five U.S. institutions, were analyzed. Physiological features were extracted from EEG, EOG, EMG, respiratory, and cardiovascular signals using standard deviation, mean absolute value, root mean square, zero-crossing rate, and Hjorth parameters (mobility and complexity). Nonlinear signal complexity was quantified using Higuchi and Petrosian fractal dimensions (HFD, PFD) computed on full-night and stage-specific EEG segments. Sleep staging dynamics were further characterized by applying fractal analysis to sleep staging sequences. To capture temporal sleep structure, sleep staging annotations along with respiratory, arousal and limb movement events were encoded as multivariate time series and projected into a low-dimensional space via incremental principal component analysis (PCA), from which latent trajectory features, including variance and fractal dimensions, were derived. All features were integrated along with demographic variables, and classified using a class-balanced Random Forest model.
Results: The proposed model achieved a 5-fold cross validation AUROC of 0.664 ± 0.069, AUPRC of 0.668 ± 0.073, accuracy of 0.601 ± 0.047, and F-measure of 0.604 ± 0.055. Performance on test dataset revealed AUROC of 0.694, AUPRC of 0.123, accuracy of 0.423, and F-measure of 0.101.
Conclusion: The proposed framework captures multi-scale sleep complexity and latent temporal dynamics, providing a promising approach for PSG-based cognitive assessment.