Predicting subject traits from EEG spectral signatures: an application to brain age

1630 

Abstract Submission 

Cecilia Jarne¹, Ben Griffin², Diego Vidaurre³

¹Aarhus University CFIN /Universidad Nacional de Quilmes/ CONICET, Aarhus, Aarhus C, ²Oxford University, Oxford, Oxford, ³Aarhus University, Aarhus, Aarhus C

Cecilia Jarne, PhD  
Aarhus University CFIN /Universidad Nacional de Quilmes/ CONICET
Aarhus, Aarhus C

Ben Griffin  
Oxford University
Oxford, Oxford

Diego Vidaurre  
Aarhus University
Aarhus, Aarhus C

We address the prediction of behavioural or cognitive traits from brain EEG spectral data. To predict from individual channels, we proposed a novel method based on interpreting the EEG spectrogram as a probability distribution. Specifically, without predefining frequency bands or relying on manual feature engineering, the proposed approach uses so-called Kernel Mean Embeddings (KME) to represent EEG spectrograms in a high-dimensional feature space (Smola et. al, 2007; Iyer et. al, 2014; Borgwardt et. al, 2006). Our method, Kernel Mean Embedding Regression (KMER), utilises Maximum Mean Discrepancy (MMD) as a distance metric on the space of probability measures, together with kernel ridge regression(Smola et. al, 2007; Saunders et al., 1998). As an example, we focus on age prediction, which can be used to explore how an individual's brain appears to have aged relative to the population average (Franke and Gaser, 2019; Smith et al, 2019. We demonstrate the effectiveness of KMER in age prediction using a multi-site resting-state EEG dataset spanning a wide age range.

This study utilises EEG data from the HarMNqEEG dataset (Li et. al, 2022), where subjects were scanned across 14 sites, to predict individuals' age using three regression methods: Ridge Regression (RR), Kernel Ridge Regression (KRR), and Kernel Mean Embedding Regression (KMER) (See Fig 1 for an illustration of how they work). In KMER, the Maximum Mean Discrepancy (MMD) is employed as a distance metric, interpreting EEG spectrograms as probability distributions. The MMD is estimated using different kernel functions to create a distance matrix. The predictive models KRR and KMER both incorporate the kernel trick for age prediction. We evaluated prediction performance using explained variance (R^2) and mean absolute error (MAE). The open-source code for this analysis is available on GitHub (Jarne et. al 2023).

KMER resulted in superior age prediction from EEG spectrograms compared to RR and KRR. KMER outperformed alternatives across all EEG channels, likely because of its capacity to capture non-linearities effectively (See Fig 2). KMER revealed that parietal sensors are the most accurate in predicting age. When predicting separately by sex, females demonstrated slightly higher prediction accuracy, suggesting a closer alignment of biological changes with chronological age. Predicting across 14 sites was harder than within site due to the wide range of age distributions. KMER, however, exhibited good accuracy and robust results even in this case.

KMER was introduced as a method for predicting individual traits from EEG spectral information. Interpreting EEG channel spectrograms as probability distributions, leverages mathematical principles from kernel learning. KMER is straightforward to implement, computationally efficient, and broadly applicable.

Despite the challenge of predicting age across sites with very different age distributions and other potential differences, the presented results showcase notable performance in age prediction in comparison to previous EEG studies. Beyond its success in age prediction from EEG data, KMER can be used in other modalities, including MEG and fMRI, and to predict other subject traits besides age. Although here we restricted ourselves to the polynomial, Gaussian, and linear kernels, the method can be further optimised through the exploration of different kernel functions and hyperparameters.

Early life, Adolescence, Aging

Classification and Predictive Modeling

EEG/MEG Modeling and Analysis ¹

Methods Development ²

Other Methods

Electroencephaolography (EEG)

Machine Learning

MEG

Modeling

Other - Kernel Methods