Multimodal Sleep Feature Extraction and Stratified Feature Selection for Dementia Prediction

Bálint Kristóf1, Szabolcs Mátyás Péter2, Kristóf Müller3, Dr. Marton Aron Goda4
1Pázmány Péter Katolikus Egyetem Faculty of Information Technology, 2Pázmány Péter Catholic University Faculty of Information Technology and Bionics, 3Pázmány Péter Catholic University - Faculty of Information Technology and Bionics, 4Faculty of Biomedical Engineering, Technion, Technion–IIT


Abstract

Cognitive impairment has been associated with alter- ations in multiple physiological processes during sleep, motivating the analysis of multimodal polysomnography beyond conventional sleep-stage measures. We devel- oped a pipeline for predicting cognitive impairment from polysomnographic recordings using modality- and sleep- stage-specific handcrafted features. The extracted fea- ture space was evaluated using missingness-based filter- ing and stratified feature selection, while additional ex- periments incorporated learned representations generated by the SleepFM foundation model. XGBoost classifiers were used for prediction. Internal experiments showed predictive information across all investigated physiologi- cal modalities and identified respiratory and Wake-stage features as particularly informative. A respiratory-only model was successfully evaluated on the hidden Challenge validation set, achieving an AUROC of 0.884, an age- weighted AUROC of 0.748, and an age-conditioned AU- ROC of 0.727. These results demonstrate that respiratory physiology alone contains substantial information asso- ciated with cognitive impairment, although the reduction after age adjustment indicates that part of this associa- tion is related to age. The internal feature analysis fur- ther suggests that informative respiratory characteristics are distributed across multiple sensors and signal descrip- tors rather than being confined to a single conventional biomarker.