Combining magnetoencephalography with magnetic resonance imaging enhances learning of surrogate-biomarkers

Denis A. Engemann, Oleh Kozynets, David Sabbagh, G Lemaître, Gaël Varoquaux, Franziskus Liem, Alexandre GramforteLife 2020 · DOI

TL;DR

This paper introduces a novel "opportunistic stacking" machine learning framework that robustly combines MEG, fMRI, and MRI data to predict brain age. This multimodal approach significantly improves prediction accuracy and reveals distinct, complementary information from each modality, even when some data is missing, making it a promising tool for clinical neuroscience.

What Problem Does This Paper Address?

Predicting health outcomes from complex brain data is a major challenge in neuroscience. While different neuroimaging modalities like MRI (structure), fMRI (activity), and electrophysiology (MEG/EEG - electrical signals) offer unique insights, it's unclear how to best combine them. Data can be redundant, common representations are hard to find, and collecting all data types from every patient can be impractical or medically contraindicated. This paper tackles how to effectively integrate these diverse data sources for robust predictive modeling, specifically using brain age as a surrogate biomarker.

Methods

The core of this work is a stacked-prediction model that combines predictions from individual modalities. This is a two-layer approach:

  • Layer 1: Ridge Regression: For each modality (MEG, fMRI, MRI), linear models (ridge regression) are trained to predict brain age. This step reduces the dimensionality of the data by creating "predictions" as new features.
  • Layer 2: Random Forest Regressor: A non-linear random forest model then takes these Layer 1 predictions as input to make a final, more accurate brain age prediction.

A key innovation is the "opportunistic stacking" approach. This allows the model to learn from any subject for whom at least one modality's data is available, rather than requiring complete data from all modalities. Missing values are handled by a special coding scheme that allows the model to potentially learn from the fact that data is missing.

Datasets

The study utilizes the Cam-CAN dataset, a large, publicly available resource comprising neuroimaging (MEG, fMRI, MRI) and neuropsychological data from 674 healthy individuals aged 17-90. This dataset is notable for its multimodal nature and large sample size, making it suitable for training complex machine learning models.

Key Insights and Ideas

  • Additive Information: MEG, fMRI, and MRI provide distinct, non-redundant information for predicting brain age. Combining them leads to significantly better predictions than using any single modality alone.
  • MEG's Contribution: The most informative MEG features for age prediction are related to cortical power spectra, specifically in the alpha (8-15 Hz) and beta (15-30 Hz) frequency bands. This suggests "spectral fingerprints" are key.
  • Opportunistic Learning is Viable: The proposed "opportunistic stacking" method effectively handles missing data, maintaining predictive performance and allowing for broader clinical application where complete datasets are rare.
  • Brain Age as a Health Indicator: The brain age prediction derived from this multimodal model correlates with cognitive and physical health markers, reinforcing its utility as a surrogate biomarker.

Key Results

  • Combining MEG, fMRI, and MRI reduced the mean absolute error (MAE) in brain age prediction to approximately 4.7 years, compared to around 6 years for MRI alone.
  • Adding MEG or fMRI individually to MRI improved prediction by about 0.8 years each.
  • The errors from fMRI and MEG predictions were only weakly correlated, suggesting they capture different aspects of brain aging.
  • MEG-based prediction was primarily driven by source power features, outperforming connectivity features.
  • The opportunistic model performed comparably to restricted models on complete data and extended performance to new cases with incomplete data.

Limitations and Future Work

  • Data Size: While Cam-CAN is large, the authors note that even larger datasets (10-100k subjects) might favor purely non-linear models over stacking.
  • Generalizability: The focus on MEG features was guided by aging literature; a broader feature set could be explored. The study also acknowledges the need for generalization testing across different sites and scanners.
  • Interpretability: While the model is predictive, detailed inferential analysis of specific feature contributions was limited to model comparisons. The authors suggest future work on inference-oriented frameworks.
  • Clinical Translation: The paper highlights the need for benchmarking and improving power topography analysis for clinical EEG, which is lower-density than MEG. Future work could also explore applying this framework to other clinical endpoints beyond age prediction.

Summary updated

Abstract

Electrophysiological methods, that is M/EEG, provide unique views into brain health. Yet, when building predictive models from brain data, it is often unclear how electrophysiology should be combined with other neuroimaging methods. Information can be redundant, useful common representations of multimodal data may not be obvious and multimodal data collection can be medically contraindicated, which reduces applicability. Here, we propose a multimodal model to robustly combine MEG, MRI and fMRI for prediction. We focus on age prediction as a surrogate biomarker in 674 subjects from the Cam-CAN dataset. Strikingly, MEG, fMRI and MRI showed additive effects supporting distinct brain-behavior associations. Moreover, the contribution of MEG was best explained by cortical power spectra between 8 and 30 Hz. Finally, we demonstrate that the model preserves benefits of stacking when some data is missing. The proposed framework, hence, enables multimodal learning for a wide range of biomarkers from diverse types of brain signals.
BibTeX
@inproceedings{engemann2020combining,
  author = {Denis A. Engemann and Oleh Kozynets and David Sabbagh and G Lemaître and Gaël Varoquaux and Franziskus Liem and Alexandre Gramfort},
  title = {{Combining magnetoencephalography with magnetic resonance imaging enhances learning of surrogate-biomarkers}},
  booktitle = {eLife},
  year = {2020},
  doi = {10.7554/elife.54055},
}
Figure