Inter-subject alignment of MEG datasets in a common representational space

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biophysics

Background:

• Pooling neural imaging data across subjects requires accurate alignment of recordings.
• Magnetoencephalography (MEG) sensor data often shows poor correlation across subjects due to anatomical and sensor location variability.
• Current alignment methods rely on source localization, which involves strong assumptions and under-determined inverse problems.

Purpose of the Study:

• To investigate an alternative method for aligning neural imaging data that bypasses source localization.
• To analyze sensor recordings directly and align their temporal signatures across subjects.
• To evaluate the robustness and performance of multiset canonical correlation analysis (M-CCA) for cross-subject data alignment.

Main Methods:

• Utilized multiset canonical correlation analysis (M-CCA), a multivariate approach, to transform individual subject data into a low-dimensional common representational space.
• Evaluated M-CCA's robustness using a synthetic dataset, analyzing the impact of noise and individual differences.
• Compared M-CCA performance against methods assuming perfect sensor correspondence and methods employing source localization on an MEG dataset.

Main Results:

• M-CCA demonstrated superior performance compared to methods assuming perfect sensor correspondence and source localization on an MEG dataset.
• The approach effectively aligns temporal signatures in sensor recordings, bypassing the need for source localization.
• Robustness was confirmed through synthetic data analysis, assessing the impact of noise and inter-subject variability.

Conclusions:

• M-CCA offers a robust and effective alternative for aligning neural imaging data across subjects, particularly for MEG.
• The method successfully aligns sensor-level data by focusing on temporal signatures, mitigating issues associated with anatomical variability and source localization assumptions.
• Potential improvements to M-CCA include incorporating spatial sensor information via a regularization term.