Magnetoencephalography Brain Spectral Biomarkers of Alzheimer's Disease: A Promising Approach Demonstrated with the BioFIND Dataset

Introduction: Magnetoencephalography (MEG) is widely used to study neurodegenerative diseases, particularly Alzheimer's disease (AD). MEG is valuable for examining functional changes in the AD brain due to its non-invasive nature and excellent spatial and temporal resolution. AD is characterized by amyloid-beta and tau protein formation, which disrupts brain networks and leads to memory and cognitive impairments. Objective: This study aimed to evaluate power spectra analysis of brain activity using various source localization methods and machine learning (ML) classification analysis based on GLMNET and correlation thresholds to identify MEG-based features that improve differentiation between healthy controls (HC) and mild cognitive impairment (MCI). Methods: We used the BioFIND dataset, including MEG data from 324 participants (158 MCI, 166 HC). MEG data were analyzed using custom MATLAB scripts with SPM12 and OSL toolboxes. Source localization methods included Bayesian LORETA (COH), Multiple Sparse Priors (MSP), Empirical Bayesian Beamformer (EBB), exact Low-Resolution Tomography Analysis (eLORETA), and Linearly Constrained Minimum Variance (LCMV). Sensor-based analyses were also conducted. Logistic Regression with L1 penalty (GLMNET) was applied in both sensor and source spaces, with different correlation thresholds for magnetometer (MAG) and gradiometer (GRAD) signals separately. We used 20 Monte Carlo iterations with 10-fold nested cross-validation to assess classifier performance. Results: Optimal performance was achieved using the LCMV method on GRAD sensors with a 0.1 correlation threshold, resulting in 74.57% accuracy, 71.55% sensitivity, and a 73.29% F1 score. Conclusions: This study demonstrates the potential of MEG as a robust tool for distinguishing between individuals with MCI and HC. By evaluating various source localization methods and correlation thresholds, we identified that the LCMV beamformer applied to GRAD signals with a 0.1 correlation threshold yielded optimal classification performance. The findings highlight GLMNET with Monte Carlo iterations and nested cross-validation as a promising framework for improving early AD detection.