Sleep Brainwaves Reveal Biological Differences in Alzheimer’s: AI Identifies Three Patient Subgroups

The pathological changes associated with Alzheimer’s disease may begin years before noticeable cognitive symptoms, but confirming these changes often relies on cerebrospinal fluid testing or imaging scans. A Spanish research team has now proposed another avenue of investigation: using machine learning to analyze electroencephalograms recorded during sleep and identify patterns of neural activity linked to disease biomarkers.

The study, published in *GeroScience*, analyzed overnight sleep data from 42 patients with mild to moderate Alzheimer’s disease and 58 people without cognitive impairment. The database was established by Hospital Universitari Santa Maria de Lleida and IRBLleida. The researchers used four EEG channels, segmented the signals by sleep stage, and then extracted linear, spectral, and nonlinear features.

The team used principal component analysis to reduce the large volume of EEG data and random forests to assess which components were more closely associated with cerebrospinal fluid markers. The resulting 30 components captured 92.4% of the variance in the original data. They not only distinguished patients from controls but also enabled a Gaussian mixture model to divide the patients into three subgroups.

These clusters were not simply ranked according to symptom severity. The three patient groups showed stepwise differences in the ratio of phosphorylated tau protein 181 to amyloid beta 42, as well as in concentrations of neurofilament light chain, a marker of axonal injury. In other words, sleep EEG may capture more than whether someone has the disease; it may also reflect different biological states among patients with Alzheimer’s disease.

Sleep provides an observational window because the neural circuits that regulate it may be affected early in the disease, with the resulting changes reflected in EEG recordings. The researchers envision that, if the method is validated in the future, more readily obtainable sleep EEG data could help identify people who need further cerebrospinal fluid testing or imaging rather than replace existing biomarkers.

The current evidence remains far from supporting clinical screening. The sample included only 100 people, all patients had already reached the mild to moderate stage, and the study did not demonstrate that the model could predict disease onset in asymptomatic or very early-stage populations. The control group also did not undergo equivalent cerebrospinal fluid marker assessments. The next step is prospective validation in larger, independent populations spanning different disease stages to establish diagnostic sensitivity, specificity, and model stability before determining whether home-based or routine sleep testing can truly serve as a reliable entry point.