Brain Age Prediction Study Based on MEG Time Series

Academic Background

With the extension of human lifespan, understanding changes in the brain throughout the life cycle has become increasingly important. The structure and function of the brain undergo significant changes with age, which not only affect cognitive functions but are also closely related to various neurodegenerative diseases (such as Alzheimer’s disease). However, current understanding of age-related changes in the brain remains incomplete, especially regarding how brain electrical activity (such as magnetoencephalography, MEG) changes with age is still unclear. To address this issue, researchers explored brain signal characteristics that can effectively predict age by analyzing large-scale resting-state MEG data from adults.

This study aims to fill gaps in existing research, particularly by identifying signal features that capture age-related changes in the brain through time series analysis techniques. These findings not only help in understanding the mechanisms of healthy aging but also provide new insights for the development of brain age prediction models.

Paper Source

This research was jointly completed by Christina Stier, Elio Balestrieri, Jana Fehring, Niels K. Focke, Andreas Wollbrink, Udo Dannlowski, and Joachim Gross. The research team comes from the University of Münster (Germany) and the University Medical Center Göttingen. The paper was published in the journal PNAS (Proceedings of the National Academy of Sciences) on February 20, 2025, titled “Temporal autocorrelation is predictive of age—an extensive MEG time-series analysis.”

Research Process

1. Data and Participants

The study used open data from the Cambridge Centre for Aging and Neuroscience (Cam-CAN). Researchers analyzed resting-state MEG data from 350 healthy adults aged between 18 and 88 years. All participants were cognitively normal individuals without major physical or mental health issues.

2. MEG Data Acquisition and Processing

MEG data was collected using a 306-channel Elekta Neuromag system with a sampling frequency of 1 kHz. The data underwent preprocessing, including the removal of external noise and correction for head movement. Subsequently, the researchers downsampled the data to 300 Hz and used independent component analysis (ICA) to remove ocular and cardiac artifacts. Finally, 30 segments of 10-second clean data were randomly selected from each participant for subsequent analysis.

3. Time Series Feature Extraction

Researchers used the Highly Comparative Time-Series Analysis (HCTSA) toolbox to extract 5,961 features from the time series of each brain region. These features include autocorrelation (AC), autoregressive coefficients, wavelet decomposition parameters, etc. Additionally, traditional frequency domain features such as specific band power, amplitude, and phase coupling were calculated.

4. Age Prediction Model

To predict the age of participants, researchers used a Partial Least Squares Regression (PLSR) model and evaluated it using 10-fold cross-validation. Model performance was assessed using Pearson’s correlation coefficient (Pearson’s r), mean absolute error (MAE), and predicted R².

Main Results

1. Predictive Performance of Traditional MEG Features

Traditional frequency domain features showed varying performance in age prediction. Amplitude and phase coupling features had lower predictive accuracy, while alpha peak frequency features performed better, especially the center of gravity method, achieving a predictive correlation of 0.65.

2. Predictive Performance of Time Series Features

Among the 5,968 time series features, 113 features had a predictive accuracy exceeding 0.7. Autocorrelation (AC) features stood out, especially at a time delay of 36 milliseconds (lag 11), where the predictive correlation reached 0.75. The prediction error of AC features was about 1.5 years lower than that of traditional features.

3. Regional Patterns of Autocorrelation Features

AC features showed significant age-related changes in the visual cortex and temporal lobe cortex of the brain. In the visual cortex, AC values increased with age, while in the temporal lobe cortex, they showed an opposite trend. These regional patterns are closely related to the aging process of the brain.

4. Predictive Performance of Multi-feature Combinations

Researchers also tested a combined model using all AC time delay features, further improving the predictive correlation to 0.82, with the mean absolute error reduced to 8.71 years. In contrast, the predictive correlation of the combined model of traditional features was 0.77.

Conclusions and Implications

This study shows that autocorrelation (AC) features can effectively capture age-related changes in the brain, especially showing significant regional patterns in the visual and temporal lobes. AC features can be used not only individually for brain age prediction but their combination model can further improve prediction accuracy. These findings provide new perspectives for understanding the mechanisms of brain aging and lay the foundation for developing more accurate brain age prediction models.

Research Highlights

1. Innovative Application of Autocorrelation Features: This study first applied autocorrelation (AC) features to brain age prediction and demonstrated their superiority in capturing age-related changes in the brain.
2. Large-scale Data Analysis: The study used MEG data from 350 adults, with a large sample size and wide age range, enhancing the universality of the results.
3. Multi-feature Combination Model: By combining multiple AC features, researchers significantly improved the accuracy of brain age prediction, demonstrating the potential of multi-feature models in brain science research.

Other Valuable Information

The study also found that the aging process of the brain exhibits different patterns in different regions, especially showing significant AC changes in the visual and temporal lobes. These region-specific changes may be closely related to cognitive decline, providing new directions for future research on the relationship between brain aging and cognitive function.

This study not only provides new tools for understanding the mechanisms of brain aging but also offers important scientific evidence for the development of brain age prediction models.