AI-driven magnetoencephalography biomarkers in dementia risk prediction: current evidence, challenges and future...

Dementia imposes a substantial global healthcare burden, with rising prevalence and limited disease-modifying treatments. Early identification of at-risk individuals is critical for timely intervention and care planning. Magnetoencephalography (MEG) provides high-temporal-resolution measurements of neuronal activity, capturing subtle functional alterations that precede clinical symptoms. Artificial intelligence (AI), particularly machine learning (ML), can leverage MEG's rich spatiotemporal information to enhance diagnostic accuracy and dementia risk prediction. This scoping review synthesizes current evidence on AI-driven MEG analysis for the classification, prediction, and prognosis of MCI and dementia, focusing on methodological approaches, predictive performance, and translational potential.

A systematic PubMed-MEDLINE search identified studies published between January 2015 and October 2025, capturing the last decade's rapid evolution of AI methodologies and their integration with neurophysiological research. Search terms combined MEG, AI, and ML with cognitive impairment and dementia. Eligible studies were peer-reviewed original research, involved human participants, employed MEG, and applied AI algorithms for classification or prediction. Extracted data included study population characteristics, MEG features, ML models, predictive biomarkers, and performance metrics.

Fourteen studies met eligibility criteria, covering populations from healthy controls to individuals with subjective cognitive decline, MCI, AD, and other dementias. MEG systems varied, with most studies employing 306-channel whole-head systems. ML algorithms ranged from traditional approaches, such as support vector machines and random forests, to deep learning architectures, including convolutional neural networks. Reported classification accuracies ranged from moderate (~60%) to high, with several studies achieving over 80% in distinguishing diagnostic categories or predicting MCI-to-AD progression. Key biomarkers included alterations in frequency-specific oscillatory activity, functional connectivity patterns, and large-scale network dynamics. Multimodal approaches integrating MEG with structural neuroimaging further improved predictive performance.

Despite heterogeneity across study designs, AI-driven MEG analyses hold significant translational potential for early, non-invasive dementia prediction, enhancing diagnostic and prognostic accuracy. Advancing clinical translation will require standardized preprocessing pipelines, larger multicenter cohorts, and explainable AI frameworks. Future research should leverage next-generation MEG technologies, such as optically pumped magnetometers, to capture brain dynamics in ecologically valid, real-world scenarios. Integrating these data with AI-driven multimodal biomarkers will improve individualized risk prediction, early diagnosis, and therapeutic decision-making in dementia.