Leveraging a Foundation Model for the EEG-Based Diagnosis of Alzheimer's Disease

Maggie Lin; Chung-Lin Hou; Tzyy-Ping Jung

Source-linked AI summary

Leveraging a Foundation Model for the EEG-Based Diagnosis of Alzheimer's Disease

Maggie Lin, Chung-Lin Hou, Tzyy-Ping Jung

arXiv:2608.27719v1Aug 27, 2026cs.LGq-bio.NC

TL;DR

Biological heterogeneity and subtle, non-stationary EEG dynamics make accurate Alzheimer’s disease diagnosis difficult for traditional hand-crafted spectral and linear methods. The study combines LaBraM embeddings with nonlinear classification, achieving strong discrimination while capturing clinically relevant spatial-spectral markers and cognitive impairment alignment.

Problem
Biological heterogeneity, low signal-to-noise ratio, and complex EEG dynamics hinder early and accurate Alzheimer’s disease diagnosis, while traditional spectral metrics may fail to distinguish early-stage patients from healthy controls.
Method
The framework extracts LaBraM embeddings from 8-second EEG epochs, pools them at the subject level, and applies nonlinear Random Forest classification alongside spectral baselines.
Results
89.36% ROC-AUC, 81.45% PR-AUC, and 82.44% Balanced Accuracy were achieved for dementia detection, with LaBraM-RF outperforming the reported spectral and linear classifier baselines.
Takeaways & Limitations
LaBraM captures clinically interpretable Alpha and Theta slowing across relevant brain regions, and higher predicted dementia probability correlates with lower MoCA scores.
Takeaways & Limitations
Sensitivity was lower than specificity, and the single-region cohort requires broader demographic and cross-device validation to establish generalizability.

Abstract

from arXiv · show

Biological heterogeneity in Alzheimer's Disease (AD) poses a critical diagnostic challenge, particularly for traditional linear methods that fail to capture non-linear neural dynamics. To address this, we propose a diagnostic framework utilizing the Large Brain Model (LaBraM), pretrained on over 2,500 hours of EEG data. By integrating these high-dimensional latent embeddings with a non-linear Random Forest classifier, our approach effectively isolates robust disease markers. Under a rigorous subject-independent 5-fold cross-validation protocol, the method achieves an ROC-AUC of 89.36% +/- 3.49%, PR AUC of 81.45% +/- 4.43%, and Balanced Accuracy of 82.44% +/- 4.34% in distinguishing dementia patients from healthy controls. Notably, this performance uses only 8-second EEG segments, surpassing traditional spectral baselines, including band-power and parameterized oscillatory features (FOOOF). Post-hoc occlusion analysis confirms the model captures clinically validated biomarkers, specifically occipital-frontal Alpha and Theta rhythm degradation. Additional neurophysiological alignment analysis demonstrated that higher LaBraM-predicted dementia probability significantly correlated with worse cognitive performance, greater clinical severity, increased theta and alpha relative power, and higher aperiodic exponent. These findings demonstrate that deep latent representations extract clinically relevant signatures from noisy signals, enabling precise, rapid, and data-efficient diagnosis.

I. INTRODUCTION

AD diagnosis is difficult because biological heterogeneity and noisy, non-stationary EEG dynamics obscure subtle prodromal signatures. The proposed framework combines pretrained LaBraM representations with nonlinear classification to capture high-dimensional disease patterns beyond conventional spectral features.

Motivation: Traditional spectral metrics yielded no statistically significant differences between early-stage AD and healthy controls.The reported bottleneck was observed for conventional metrics, including Relative Band-Power and Alpha-to-Theta ratio.
Motivation: LaBraM uses self-supervised pretraining on over 2,500 hours of unlabeled EEG data to learn context-aware neural representations.These representations are intended to capture generalized spatiotemporal features inaccessible to conventional spectral decomposition.
Proposed approach: The study integrates pretrained LaBraM embeddings with a nonlinear Random Forest classifier for AD detection.The pipeline explicitly tests whether AD-related biomarkers are better separated by nonlinear decision boundaries in the latent space.
Proposed approach: The framework benchmarks LaBraM-Random Forest against traditional Band-Power and FOOOF spectral baselines.This comparison evaluates foundation-model embeddings alongside interpretable and parameterized spectral representations.

A. Proposed Methodological Framework

The proposed framework tokenizes EEG epochs, transforms them through pretrained LaBraM embeddings, and classifies the resulting latent representations with Random Forest. It benchmarks this pipeline against traditional Band-Power and FOOOF feature paradigms.

A. Proposed Methodological Framework: The study compares traditional Band-Power features, FOOOF parameters, and pretrained LaBraM embeddings.These three representation paradigms form the comparison between interpretable spectral features and foundation-model representations.
LaBraM representation: EEG signals are segmented and tokenized into patches before projection into a high-dimensional latent space with temporal and spatial positional encodings.The encodings preserve chronological order and electrode-array topology.
LaBraM representation: LaBraM’s Transformer encoder applies self-attention and feed-forward layers to capture long-range dependencies and nonlinear interactions.Its output is a context-aware feature vector representing latent brain-signal dynamics.
Downstream classification: Random Forest classifies LaBraM embeddings through nonlinear partitioning of the latent EEG space.Bootstrapped samples, random feature subsets, and ensemble averaging support high-dimensional feature handling while reducing overfitting.

3) Dimensionality Reduction via Principal Component Analysis (PCA):

PCA with whitening compresses high-dimensional embeddings before classification, while SVM settings are selected on validation data and evaluated on an isolated test set.

3) Dimensionality Reduction via Principal Component Analysis (PCA):: PCA with whitening reduces the 200-dimensional embedding to 44 dimensions while retaining 95% of explained variance.This is a 78% dimensionality reduction intended to filter noise while preserving pathological signatures.
3) Dimensionality Reduction via Principal Component Analysis (PCA):: The whitening transformation isotropizes feature covariance and is especially relevant to optimizing the SVM decision boundary.The study checks that this transformation does not remove variance needed by Random Forest partitioning.
SVM configuration: The SVM objective weights errors inversely to class frequency so dementia misclassifications receive greater penalty than control misclassifications.This weighting addresses class imbalance in the classification task.
SVM configuration: The regularization strength C and RBF coefficient γ are optimized on an independent validation set using balanced accuracy.The selected model is then evaluated on a strictly isolated test set for unbiased metrics.
SVM configuration: The RBF kernel is used for the SVM classification task.The accompanying description identifies x_i and x_j as input feature vectors and γ as the parameter controlling individual-sample influence.

B. Baseline Feature Extraction for Comparison

The comparison uses canonical PSD band-power features and FOOOF’s periodic/aperiodic parameterization, evaluated with the same classifiers and subject-independent cross-validation. These baselines characterize disease-related shifts toward theta activity and away from beta activity.

B. Baseline Feature Extraction for Comparison: PSD baselines summarize EEG across Delta, Theta, Alpha, Beta, and Gamma frequency ranges.They are designed to capture increased low-frequency activity and decreased higher-frequency oscillatory activity.
B. Baseline Feature Extraction for Comparison: FOOOF decomposes spectra into aperiodic offset and exponent plus periodic peak frequency, power, and bandwidth.The model is fit over 3–40 Hz using BIC to provide a granular representation while limiting overfitting.
B. Baseline Feature Extraction for Comparison: Absolute power increases in the theta band as disease severity progresses, while relative spectra shift from alpha-dominant activity in healthy subjects toward theta dominance in dementia.Theta-dominance maps show progressively stronger theta dominance across Healthy, SCD-MCI, and Dementia groups.
B. Baseline Feature Extraction for Comparison: Dementia patients show increased theta power and reduced beta power in relative band-power comparisons.The figure caption reports significant group differences in these bands.
B. Baseline Feature Extraction for Comparison: Band-Power and FOOOF baselines use Random Forest and RBF-SVM under the same subject-independent 5-fold cross-validation framework.Results are reported as mean ± standard deviation for a consistent comparison.

C. Experimental Setup and Data Preprocessing

The study used multi-center resting-state EEG from 308 participants, with clinical assessments defining diagnostic ground truth. For binary Control–Dementia classification, MCI participants were excluded, leaving 206 subjects.

C. Experimental Setup and Data Preprocessing: 308 participants contributed 30-channel, 120-second eyes-open resting-state EEG recorded across four medical centers in Taiwan during 2021–2022.The cohort included Control, MCI, and Dementia groups, with clinical diagnoses serving as ground truth.
C. Experimental Setup and Data Preprocessing: Clinical characterization included MMSE, MoCA, CDR, GDS, IADL, and ADL assessments.
C. Experimental Setup and Data Preprocessing: 206 subjects remained for binary Control–Dementia classification after excluding 102 MCI patients.The dataset was partitioned into training, validation, and independent testing sets.
C. Experimental Setup and Data Preprocessing: The fixed data split supported hyperparameter tuning, while subject-level pooled embeddings were evaluated with 5-fold cross-validation.

2) Data Preprocessing:

EEG preprocessing removed artifacts, standardized signals, and extracted both conventional spectral features and LaBraM embeddings from short epochs. Subject-level pooling and subject-independent 5-fold evaluation were used to avoid treating epochs as independent subjects.

2) Data Preprocessing:: Raw 500-Hz EEG was filtered from 0.1–50 Hz, downsampled to 200 Hz, artifact-cleaned, re-referenced with CAR, and Z-score normalized.Frontopolar channels were reconstructed with spherical spline interpolation, and ICA removed ocular, myogenic, and cardiac artifacts.
2) Data Preprocessing:: PSD, absolute and relative band-power, and theta-dominance features were computed across canonical EEG bands before group-level testing.Relative power was normalized by total 1–45 Hz power.
2) Data Preprocessing:: FOOOF decomposed spectra into periodic and aperiodic components, while MCI data were retained only for exploratory spectral analysis.MCI groups were excluded from downstream binary classification.
2) Data Preprocessing:: Clean EEG was segmented into 8-second epochs, with 75% overlap applied only to Dementia recordings and non-overlapping windows used for Controls.LaBraM embeddings were extracted per epoch, pooled by subject, and classified using Random Forest, Logistic Regression, and SVM models.
2) Data Preprocessing:: Performance used subject-level pooled embeddings in 5-fold cross-validation, with each fold training on four subject partitions and testing on one.Metrics were reported as mean ± standard deviation, and baseline comparisons used one-tailed paired t-tests on ROC-AUC scores.

6) Clinical Interpretability:

Interpretability analyses linked model behavior to spectral background changes and anatomical EEG regions. LaBraM features also showed stronger classification performance than Band-Power and FOOOF features.

6) Clinical Interpretability:: Band-stop occlusion across Delta, Theta, Alpha, Beta, and Gamma bands quantified each band’s contribution through ROC-AUC decline.Single-channel occlusion similarly identified the 10 most influential sensors among 30 channels.
6) Clinical Interpretability:: Dementia spectra showed increased theta activity and significantly reduced beta power, consistent with EEG slowing.
6) Clinical Interpretability:: Theta-band differences were primarily associated with aperiodic offset changes rather than periodic theta peak power.The reported group difference was significant for aperiodic offset (p = 0.0059) but not periodic theta peak power (p = 0.372).
6) Clinical Interpretability:: LaBraM features achieved an AUC of 89.4% ± 3.5%, compared with 65.1% ± 11.3% for Band-Power and 59.9% ± 7.8% for FOOOF.

2) Classifier Performance:

Across subject-independent 5-fold cross-validation, LaBraM-Random Forest produced the strongest reported classification performance and a balanced confusion-matrix profile. Its improvements over spectral baselines were statistically significant.

2) Classifier Performance:: 89.4% ± 3.5% ROC-AUC was achieved by LaBraM-Random Forest, exceeding PCA-SVM-RBF at 81.7% ± 5.4% and SVM-RBF at 82.2% ± 10.5%.
2) Classifier Performance:: Linear models performed near chance, with Logistic Regression at 48.6% ± 5.3% AUC and Linear SVM at 47.3% ± 6.7% AUC.
2) Classifier Performance:: 85.7% specificity was obtained for Healthy controls, alongside 78.8% sensitivity for Dementia patients.These values came from the aggregated confusion matrix across five validation folds.
2) Classifier Performance:: The proposed framework significantly outperformed Band-Power with Random Forest (p = 0.004) and FOOOF with SVM-RBF (p = 0.009).The tests used one-tailed paired t-tests on five cross-validation AUC scores.

3) Statistical Benchmarking:

Nonlinear decision boundaries better capture the dementia-related structure in LaBraM representations than linear alternatives, while predicted dementia probability aligns with clinical and neurophysiological impairment.

Linear versus Non-Linear Classifier:: Manifold accuracy was 54% for linear classifiers versus 79% for the RBF kernel, with a significant performance gap (p < 0.01).The comparison supports nonlinear separation of pathological heterogeneity.
Linear versus Non-Linear Classifier:: UMAP visualizations compare Linear, LinearSVC, Polynomial, and RBF decision boundaries in a reduced two-dimensional embedding space.The UMAP settings were n neighbors = 10 and min dist = 0.1.
Neurophysiological Alignment:: Higher LaBraM-predicted dementia probability was associated with lower MoCA scores and greater clinical severity.These associations indicate alignment with cognitive impairment.
Neurophysiological Alignment:: Higher predicted dementia probability was also associated with increased theta-alpha relative power and higher aperiodic exponent after false discovery rate correction.These EEG features provide interpretable neurophysiological correlates of model output.

6) Post-hoc analysis of Influential Brain Regions and Frequency Bands:

Occlusion analysis identifies Alpha and Theta rhythms and specific frontal, occipital, and central channels as influential features in the LaBraM-RF pipeline.

Spectral Importance:: Occluding Alpha (8–13 Hz) and Theta (4–8 Hz) reduced ROC AUC by 8.91% and 7.85%, respectively.These were the most critical frequency bands in the spectral perturbation analysis.
Spatial Importance:: C4 and O1 were the most influential channels, with removal of either producing a 7.45% performance drop.Other highly ranked sensors were F8, O2, and Fp1.
Spatial Importance:: Predictive power concentrated in occipital and frontal regions, including O1, O2, Fp1, and F8.The spatial pattern aligns with alpha rhythm degradation and frontal slowing described for Alzheimer’s disease.

IV. CONCLUSION AND DISCUSSION

The study combines LaBraM embeddings with nonlinear Random Forest classification for dementia detection and identifies clinically interpretable spectral-spatial markers. Performance remained strong despite a heterogeneous control group, but sensitivity was lower than specificity and broader validation is needed.

Conclusion and Discussion: LaBraM with a nonlinear Random Forest achieved ROC-AUC 89.36%, Balanced Accuracy 82.44%, and PR-AUC 81.45%.The framework uses standard EEG and targets complex nonlinear pathological signatures.
Conclusion and Discussion: Alpha and Theta reliance indicates that LaBraM captures spectral slowing associated with Alzheimer’s disease.Higher predicted dementia probability also correlated with lower MoCA scores, reflecting cognitive impairment.
Conclusion and Discussion: Influential channels included occipital O1 and O2, frontal Fp1 and F8, central C4, and temporal P7.These regions were linked to alpha degradation, frontal slowing, and networks involved in sensory-visual integration, executive function, and memory.
Limitations: The heterogeneous control group included individuals with Subjective Cognitive Decline, which may introduce label noise.The reported specificity remained 85.7% despite this limitation.
Limitations: Sensitivity was 78.8%, lower than specificity at 85.7%, and the single-region cohort limits demonstrated generalizability.Future validation should include broader demographic diversity and cross-device hardware testing.
Future Directions: Future work will extend the framework to Mild Cognitive Impairment to study prodromal neurodegenerative trajectories.The stated goal is to assess whether latent topological features can discriminate transitional MCI signals.