Towards improving Alzheimer's intervention: a machine learning approach for biomarker detection through combining MEG and MRI pipelines

OBJECTIVE: To identify robust biomarkers for early cognitive decline, classification performance was evaluated for healthy controls (HC) versus mild cognitive impairment (MCI) using MEG-derived spectral features at the sensor and source levels, and by testing whether adding MRI anatomical features improves discrimination in the BioFIND cohort.

APPROACH: Data from 324 participants (166 HC and 158 MCI) were analyzed using automated artifact rejection and source reconstruction. Spectral biomarkers, including sensor-space and sourcespace representations, were extracted from the magnetometer and gradiometer signals. Five inverse-solution methods were assessed (including linearly constrained minimum variance [LCMV] and exact low-resolution electromagnetic tomography analysis [eLORETA]) to increase the robustness of source-based biomarker selection. Multiple classifiers, including a support vector machine (SVM) and L1-regularized logistic regression (GLMNET), were compared using nested K-fold cross-validation embedded in Monte Carlo replications to limit overfitting.

MAIN RESULTS: The best performance was obtained by combining MRI features with source-based MEG features using GLMNET, achieving 76.31 ± 1.47% accuracy and 82.44 ± 1.44% AUC. MEG-only models showed strong discrimination and outperformed MRI-only analyses, particularly those based on the LCMV and eLORETA features. Across settings, the SVM and GLMNET were the most consistently effective classifiers.

SIGNIFICANCE: These findings support MEG/EEG spectral biomarkers, particularly source-based representations, as sensitive markers of early AD-related functional changes. Integrating MEG with MRI can further improve HC-versus-MCI classification. The study also demonstrated that rigorous validation (nested cross-validation with Monte Carlo replication) is essential for reliable biomarker assessment and for comparing preprocessing, inverse-solution, and classifier choices in translational neuroimaging pipelines.