Integrating Sleep Macroarchitecture and Multi-Modal PSG Features for Cognitive Impairment Prediction

Parisa Sattar1, Marco Orrù2, Nicla Mandas3, Eduardo Caballero Saldivar4, Giulia Baldazzi5, Danilo Pani6
1MeDSP Lab, Department of Electrical and Electronic Engineering, University of Cagliari, Italy; Interdepartmental Sleep Disorder Research Center, University of Cagliari, Cagliari, Italy;, 2DIBRIS, University of Genova; MeDSP Lab, DIEE, University of Cagliari, 3The Hadron Academy, IUSS, Pavia; MeDSP Lab, DIEE, University of Cagliari, 4Aragon Institute of Engineering Research, University of Zaragoza, 5MeDSP Lab, DIEE, University of Cagliari;, 6MeDSP Lab, Dept. of Electrical and Electronic Engineering, University of Cagliari

Abstract

Cognitive impairment (CI) represents a growing public health burden, and early identification before clinical manifestation remains a key challenge. Polysomnography (PSG) provides a multimodal assessment of sleep, and may therefore capture early polypathic dysfunction preceding CI diagnosis. To this aim, for the George B. Moody PhysioNet Challenge 2026, MeDSP team proposes a multi-domain feature-based framework for predicting future CI from overnight PSG recordings. Leveraging the Human Sleep Project database of 6600 PSG recordings, several features spanning sleep architecture, EEG spectral content, oxygen saturation, respiratory events, arousals, limb movements, and hypnogram complexity were extracted. Among them, the 20 highest-scoring features were selected and passed to a 300-tree Random Forest classifier. Model performance was preliminarily assessed via leave-one-site-out (LOSO) cross-validation on the training set, in terms of age-conditioned AUC (AC-AUC) and standard metrics. Across the LOSO runs, the proposed framework reached an AC-AUC of 0.68±0.04. On the hidden validation set, the model obtained an AC-AUC of 0.708 and a Challenge reward of 0.12, indicating its potential as an explainable tool for the early prediction of CI.