Quantitative sleep EEG identifies CSF core biomarker-related subgroups in Alzheimer's disease

Early detection and biological characterization of Alzheimer's disease (AD) remain challenging, as current diagnostic approaches rely on invasive cerebrospinal fluid (CSF) sampling or costly neuroimaging, limiting scalability. Sleep quantitative electroencephalography (qEEG) provides a non-invasive measure of brain function and may capture early AD-related neural alterations; however, the high dimensionality and complexity of these features limit interpretation with conventional approaches, requiring multivariate methods. The objective of this study is to assess whether sleep qEEG features are associated with biologically meaningful stratification of AD in accordance with the NIA-AA 2024 framework. Forty-two patients with mild-to-moderate AD underwent overnight polysomnography and CSF biomarker assessment, while 58 cognitively unimpaired controls provided sleep EEG recording. EEG signals from four channels were preprocessed, segmented by sleep stage, and characterized using linear, spectral, and non-linear features. Dimensionality reduction was performed using principal component analysis (PCA), guided by random forest-based relevance to CSF biomarkers (A $\beta$ 42, p-tau181, t-tau, and neurofilament light chain (NfL). Gaussian mixture models (GMMs) were applied to patient-level representations, including derived hybrid ratios (p-tau181/A $\beta$ 42), to identify biologically coherent subgroups. A reduced 30-component qEEG representation explaining 92.4% of the variance differentiated cognitively unimpaired individuals from patients with AD and identified three AD subclusters. These subgroups showed graded differences in CSF biomarker profiles, including p-tau181/A $\beta$ 42 and NfL. Sleep qEEG features show structured associations with CSF biomarker profiles and capture variability across the AD continuum. These findings suggest that higher-level qEEG-based machine learning approaches may complement established biomarker-based methods for biological characterization of AD within a geroscience framework.