TL;DR
This study demonstrates that low-cost, at-home mobile EEG devices can accurately predict a person's age, achieving performance comparable to lab-grade equipment. By analyzing over 5200 individuals' sleep and meditation recordings, the researchers found that sleep data yielded better age prediction, with specific sleep stages (N2-N3) containing the most informative signals.
What Problem Does This Paper Address?
Electroencephalography (EEG) is a powerful tool for understanding brain activity, but its use has traditionally been confined to controlled laboratory settings. This limits its accessibility and the ability to collect large-scale, real-world data. While advancements in sensor technology have led to affordable, wearable EEG devices, questions remain about the impact of reduced spatial resolution and increased noise on the reliability of brain assessments. This paper tackles the challenge of using these low-cost, at-home EEG devices for biomarker discovery, specifically focusing on the "brain age" framework. Brain age, the difference between predicted and chronological age, is a promising indicator of brain health and potential pathology. However, its application has been largely restricted to expensive MRI-based studies, hindering its widespread use in everyday settings.
Methods
The researchers employed a machine-learning pipeline for age prediction, building upon established methods for brain age analysis. Key steps included:
- Preprocessing: Standard EEG preprocessing techniques were applied, including zero-phase FIR band-pass filtering (0.1-49 Hz), notch filtering to remove hardware-specific noise and power line interference, and downsampling to 128 Hz.
- Feature Extraction: The core of their approach involved extracting filterbank covariance matrices. This method captures the relationships between different brain regions (spatial) and across various frequency bands. They explored three types of representations:
- Spectral: Log-powers of each channel in each frequency band.
- Spectro-spatial: Covariability of channel pairs within each frequency band.
- Cross-spectro-spatial: Capturing interactions between channels across different frequency bands.
- Vectorization: These covariance matrices were then vectorized into feature vectors. They used two approaches: log-diagonal extraction and Riemannian tangent space projection.
- Machine Learning Model: A Ridge regression model, a common choice for EEG-based regression tasks, was used to predict age.
- Sleep-Specific Vectorization: For sleep data, they further refined the process by extracting features for each sleep stage (W, N1, N2, N3, R) separately, allowing the model to learn stage-specific patterns.
- Evaluation: Models were evaluated using Monte Carlo cross-validation, measuring performance with the coefficient of determination ($R^2$) and mean absolute error (MAE). Longitudinal data was used to assess the stability of brain age predictions over time.
Datasets
The study utilized data from InteraXon's anonymized database of Muse S users, a wireless, dry EEG device. They curated four datasets:
- Muse Meditation Dataset (MMD): 4191 meditation recordings from unique individuals.
- At-home Muse Unlabelled Sleep Dataset (AMUSED): 1020 overnight sleep recordings from unique individuals.
- Longitudinal Muse Meditation Dataset (MMD-long): 497 meditation recordings from four subjects with multiple sessions.
- Longitudinal At-home Muse Unlabelled Sleep Dataset (AMUSED-long): 539 sleep recordings from four subjects with multiple sessions.
These datasets represent a significant scale for at-home EEG studies, with over 5200 unique individuals contributing data.
Key Insights and Ideas
- At-home EEG is viable for brain age prediction: The study convincingly shows that low-cost, consumer-grade EEG devices can yield age prediction performance comparable to lab-grade equipment, opening doors for widespread accessibility.
- Sleep data is richer for age prediction: Overnight sleep recordings consistently outperformed meditation recordings in predicting age, suggesting that the dynamic and complex nature of sleep provides more age-related information.
- Specific sleep stages matter most: Within sleep, the N2 and N3 stages were found to contain the most crucial signals for age prediction, aligning with known changes in brain activity during these deep sleep phases.
- Multivariate features are key: Combining spectral, spatial, and cross-spectral information significantly improved age prediction accuracy, highlighting that age-related brain patterns are distributed across different brain regions and frequency interactions.
- Brain age captures both trait and state information: Longitudinal analysis revealed that while brain age predictions are generally stable over longer periods (trait-like), there's also variability that reflects short-term fluctuations (state-like), offering a nuanced view of brain health.
Key Results
- Meditation: Models achieved a median $R^2$ of 0.34 (MAE = 9.27 years), significantly above chance.
- Sleep: Models achieved a higher median $R^2$ of 0.47 (MAE = 7.57 years) when using all sleep stages.
- N2 and N3 stages individually yielded $R^2$ of 0.37.
- W and R stages showed lower performance ($R^2$ of 0.20-0.22).
- Feature Importance: Spectral features alone provided above-chance performance. Adding spatial and cross-spectral information further improved accuracy. For meditation, low and mid alpha bands were most important. For sleep, alpha in N2 and delta/alpha in N3 were critical.
- Longitudinal Stability: Despite day-to-day variability, average brain age predictions remained stable over months for individual subjects, with within-subject standard deviations in brain age delta being lower than cross-subject standard deviations.
Limitations and Future Work
- Uncontrolled Environment: The data was collected at home without expert supervision, meaning no strict adherence to recording instructions could be guaranteed. This contrasts with most existing brain age studies that use clinical or research datasets.
- Limited Metadata: Minimal metadata was available, preventing the researchers from knowing if participants had medical conditions. This could potentially lower the sensitivity of their models to pathology-related EEG activity compared to studies trained on healthy individuals.
- Lack of Clinical Validation: The absence of clinical cognitive assessments limited the ability to explore the semantic meaning of brain age predictions (e.g., linking positive brain age to cognitive slowing).
- Future Work: The authors suggest that future research could explore combining information from multiple awake states to potentially bridge the performance gap between meditation and sleep-based age prediction. Further investigation into the precise connection between anatomical and functional measures could also be beneficial. Additionally, head-to-head comparisons between different mobile EEG devices would be valuable.
Summary updated