Abstract
Magnetoencephalography (MEG) is widely used to study neurodegenerative disorders, particularly Alzheimer's disease (AD). AD is linked to amyloid-beta and tau protein formation, which disrupts brain anatomical and functional networks, leading to memory and cognitive impairments. Due to its noninvasive nature and excellent temporal resolution, MEG is valuable for examining functional changes in the AD brain. We evaluated a pipeline, based on nested cross-validation with Monte-Carlo replications, with MEG-derived sensors and source-based spectral features to discriminate healthy controls (HC) versus mild cognitive impairment (MCI) brain activity features. We also compared the effectiveness of combining MEG and MRI features extracted for 324 participants (158 MCI, 166 HC) in the BioFIND dataset. A robust selection of brain source activity biomarkers was implemented through five independently tested inverse solutions, including a linearly constrained minimum variance (LCMV) beamformer and exact low-resolution electromagnetic tomography (eLORETA). Several machine learning classifiers were also evaluated, including Support Vector Machine (SVM) and Logistic Regression with L1 penalty (GLMNET). Initial results showed that combining MRI features with source-based MEG features yielded the best performance based exclusively on spectral features (Acc=76.31±1.47%) using the GLMNET classifier. MEG features alone, particularly those extracted from LCMV and eLORETA analyses, demonstrated good performance (Acc=74.77±1.57%), surpassing MRI- and sensor-based analyses using an SVM classifier (Acc=72.74±1.34% and Acc=69.29±1.68% respectively). However, ongoing more advanced analyses relying on features extracted from LCMV and eLORETA and derived functional connectivity solutions (coherence - COH, imaginary COH - iCOH, wPLI, AEC, EIC) have surpassed these and previous benchmarks for the BioFIND dataset (Acc=94.91±0.01%, F1=94.63±0.01%, MCC=94.96±0.01% for MEGMAG-LCMV-wPLI; Acc=93.29±0.01%, F1=93.14±0.01%, MCC=93.30±0.01% for MEGMAG-LCMV-EIC; Acc=92.16±0.01%, F1=92.16±0.01%, MCC=93.30±0.01% for MEGGRAD-LCMV-EIC), while using a similar pipeline. Our findings highlight the potential of using MEG signals to find MCI biomarkers, supporting critical early intervention. They also provide evidence on potential "best" approaches to analysing magneto-electroencephalography-based brain activity.
Original language | English |
---|---|
Publication status | Published online - 30 Oct 2024 |
Event | MEG-UKI 2024 - University of Birmingham, Birmingham, United Kingdom Duration: 30 Oct 2024 → 31 Oct 2024 https://uobevents.eventsair.com/meguki/ |
Conference
Conference | MEG-UKI 2024 |
---|---|
Country/Territory | United Kingdom |
City | Birmingham |
Period | 30/10/24 → 31/10/24 |
Internet address |
Keywords
- MEG
- BioFIND dataset
- Alzheimer disease
- mild cognitive impairment
- inverse problem
- functional brain connectivity
- machine learning
Fingerprint
Dive into the research topics of 'Magnetoencephalography brain biomarkers of Alzheimer's disease: an ongoing study based on the BioFIND dataset'. Together they form a unique fingerprint.Prizes
-
Overall Best Paper Award given by IEEE-EMBS Conference on Biomedical Engineering and Sciences
Ahmad, A. L. (Recipient), Faye, I. (Recipient), Sanchez Bornot, J. (Recipient), Idris, Z. (Recipient), Coyle, D. (Recipient) & Sotero, R. C. (Recipient), 17 Dec 2024
Prize
File