Predicting Subject Traits From Brain Spectral Signatures: An Application to Brain Ageing

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Predicting subject traits from brain data is crucial for clinical research and understanding cognition.
• Electroencephalography (EEG) offers a non-invasive, cost-effective alternative to neuroimaging for brain data acquisition.
• Manual feature extraction from complex EEG data can introduce bias and limit prediction accuracy.

Purpose of the Study:

• To investigate data-driven Kernel methods for trait prediction using single-channel EEG spectrograms.
• To develop a method that avoids manual feature extraction by treating EEG spectrograms as probability distributions.
• To compare the performance of Kernel mean embedding regression against traditional Kernel ridge regression and non-Kernelized approaches.

Main Methods:

• Reinterpreting single-channel EEG spectrograms as probability distributions.
• Applying Kernel mean embedding regression for trait prediction.
• Comparing Kernel methods with non-Kernelized approaches on EEG data.

Main Results:

• Kernel methods demonstrated improved prediction performance compared to non-Kernelized methods.
• The capacity of Kernel methods to handle nonlinearities between EEG spectrograms and traits was key to improved performance.
• The developed method successfully predicted biological age in a multinational EEG dataset (HarMNqEEG), showing generalization capabilities.

Conclusions:

• Data-driven Kernel methods offer a robust and accurate approach for trait prediction from EEG data.
• Treating EEG spectrograms as probability distributions enables advanced machine learning without manual feature engineering.
• The Kernel mean embedding regression technique shows promise for diverse applications in neuroscience and psychology.