MEG Spectral Biomarkers of Alzheimer’s Disease: Integrating MEG and MRI Features Using the BioFIND Dataset

Background: Alzheimer’s disease (AD), the leading cause of dementia, causes brain communication disruption due to the gradual accumulation of amyloid-beta and tau protein, which can be associated with behavioural changes linked to memory and cognition impairment. Magnetoencephalography (MEG) and magnetic resonance imaging (MRI) offer non-invasive ways to study these brain changes. Classification analysis combining MEG power spectral density with MRI features can provide critical biomarker information, enabling discrimination of healthy control (HC) individuals from individuals with mild cognitive impairment (MCI).

Objective: This study has a two-fold goal: assessing the importance of MEG power spectrum cortical features combined with MRI brain regional volumetric information, and the evaluation of a novel classification approach based on the sign-constrained logistic regression with L1 regularization (called GLMNET for simplicity, as it is implemented using the R’s GLMNET package).

Methods: We utilized the BioFIND dataset, which includes MEG recordings from 324 participants (158 MCI and 166 HC), accompanied by MRI data for nearly all individuals. MEG data were processed using custom MATLAB scripts based on the SPM12 and OSL toolboxes. MEG-based source localization was performed using mainly two methods: linearly constrained minimum variance (LCMV) beamforming and exact low-resolution electromagnetic tomography (eLORETA), which were applied separately for MEG’s magnetometer (MAG) and gradiometer (GRAD) signals. Freesurfer was utilized to extract brain regional volumetric information. Correlation analysis was performed a priori between features and labels, exclusively for the training subset, to preselect “important” features and extract sign information for different thresholds (0:0.05:0.25), and this information was utilized during the GLMNET classification analysis. Finally, we implemented a classification pipeline based on Monte Carlo replication of a 10-fold nested cross-validation analysis to perform a robust assessment of performance metrics, including accuracy and F1 score.

Results: Highest performance was obtained by combining all the extracted LCMV MAG and GRAD features and MRI features using 0-correlation threshold, with an accuracy of 77.9% and F1 score of 75.7%.

Conclusions: This study demonstrates the effectiveness of MEG in distinguishing healthy aging from cognitive decline. By comparing classification performance for different biomarkers, as selected through multiple source localization methods and varying correlation thresholds, we assessed their possible importance more robustly. Moreover, this decision took into consideration that brain inverse solution approaches follow different assumptions and using that to our advantage together with exploiting sign information can enhance biomarker interpretation.