Wearable-Compatible Detection of Mild Cognitive Impairment Using Novel Features Based on Sleep Stage Dynamics

Background: Mild Cognitive Impairment (MCI) is an early stage of cognitive decline and a major risk factor for dementia, typically diagnosed using neuropsychological assessments such as memory and executive function tests. While EEG-based detection has been widely explored, most approaches rely on raw signal analysis and computationally intensive deep learning models. In contrast, wearable devices use indirect behavioral proxies (e.g., activity patterns or sleep–wake patterns), limiting diagnostic specificity. Although substantial clinical evidence indicates altered sleep architecture in MCI, the use of sleep stage dynamics for MCI classification remains largely unexplored. Methods: We propose a lightweight and physiologically interpretable framework using novel features based on hypnogram-derived sleep dynamics. The method was evaluated on the MASS SS1 dataset (36 healthy, 17 MCI subjects) using five classifiers—Logistic Regression, Random Forest, XGBoost, Linear SVM, and RBF SVM—with leave-one-subject-out validation and threshold optimization. Results: RBF SVM achieved the best performance (accuracy: 77.4%, balanced accuracy: 78.7%, sensitivity: 82.4%, specificity: 75.0%, ROC AUC: 0.778), followed by Random Forest (accuracy: 77.4%, balanced accuracy: 77.1%) and XGBoost (accuracy: 71.7%, balanced accuracy: 73.0%). Conclusions: This proof-of-concept study demonstrates that features extracted from sleep stage dynamics are effective, non-invasive, and interpretable biomarkers for early MCI detection, with strong potential for integration into wearable sleep monitoring systems.