What question did this study set out to answer?

This research aims to evaluate the predictive capacity of EEG features for early-stage Alzheimer's disease and their correlation with cognitive function and CSF biomarkers.

June 1, 2026Open Access

EEG biomarkers can predict early-stage Alzheimer’s disease and correlate with intracerebral pathology: a multimodal machine learning study

Key Points

This research aims to evaluate the predictive capacity of EEG features for early-stage Alzheimer's disease and their correlation with cognitive function and CSF biomarkers.
Included 101 patients with mild cognitive impairment and 69 healthy controls matched for age and education.
Utilized five machine learning algorithms to analyze EEG data, focusing on power spectral density and microstates analysis.
Conducted correlation and mediation analysis to explore relationships between EEG features and cognitive metrics.
Logistic Regression model achieved the highest training performance (mean AUC = 0.859 ± 0.059).
Test set showed optimal predictive performance using PSD and microstates features (AUC = 0.949, 95% CI: 0.877-1.000).
Negative correlation observed between theta frequency band and CSF Aβ₁₋₄₂ levels (r = -0.31, p = 0.015), indicating mediation effect on cognitive scores.

Abstract

BACKGROUND: Early recognition of Alzheimer's disease (AD) is crucial for timely intervention and delaying disease progression. Electroencephalogram (EEG) technology provides a direct reflection of the brain's dynamic activity. However, the relationship between potential EEG features and cognitive function in early-stage AD patients, as well as cerebrospinal fluid (CSF) pathological biomarkers, remains unclear. METHODS: This study included 101 patients with mild cognitive impairment (MCI) and mild AD, alongside 69 healthy controls (HC) matched for gender, age, and educational attainment. Extracting EEG power spectral density (PSD) and microstates analysis features as training features for machine learning (ML), we employed five ML algorithms-Support Vector Machines (SVM), Logistic Regression (LR), Random Forests (RF), XGBoost, and LightGBM-for training and testing. Model performance was assessed using the area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations (SHAP) plots were employed to elucidate variable importance within the model, and sequential forward selection (SFS) was utilised to identify potential features. Correlation analysis and mediation analysis were conducted to investigate the relationships between EEG features, CSF pathological biomarkers, and cognitive function. RESULTS: The LR model demonstrated the highest average predictive performance in the training set (mean AUC = 0.859 ± 0.059). The model incorporating PSD and microstates features demonstrated optimal predictive performance in the test set (AUC = 0.949, 95% CI: 0.877-1.000), outperforming any single-feature model. Based on SHAP and SFS analyses, six potential EEG indicators were identified: central region delta frequency band, central region theta frequency band, temporal region beta frequency band, microstate mean duration, microstate C duration, and the transition probability from microstate C to A. Mediation analysis revealed a significant negative correlation between central region theta frequency band and CSF Aβ₁₋₄₂ levels (r = - 0.31, p = 0.015), and the central region theta frequency band mediated the relationship between Aβ₁₋₄₂ levels and Mini-Mental State Examination (MMSE) scores (indirect effect = 0.0007, 95% CI: 0.0001-0.0013). CONCLUSION: The combined application of EEG and ML enables efficient classification diagnosis of early-stage AD, and EEG is correlated with intracerebral pathological biomarkers and cognitive impairment.

Mark Helpful

Bookmark

Relay

View Full Paper