Predicting subject traits from M/EEG spectrograms using kernel mean embedding

CECILIA JARNE; BEN GRIFFIN; DIEGO VIDAURRE

Dublin

Conferencia; MEG-UK; 2023

Trinity Collage

Predicting subject traits from data is crucial for understanding cognitive processes, brain disorders, diseases and normal ageing. In this work, we propose a mathematically principled method to predict subject traits from M/EEG spectrograms. The idea is to interpret a spectrogram as a probability distribution and apply the Kernel Mean Embedding of distributions, a powerful kernel-based approach that takes probability distributions as inputs. Focusing on both accuracy and robustness, we demonstrate its use and improvement for age estimation over a baseline method like ridge regression by leveraging the HarMNqEEG dataset—a multinational compilation of EEG recordings—, which we assessed using leave-one-country-out cross-validation. Our method shows key insights, such as identifying brain regions with larger effects on ageing and intriguing gender-related patterns. We observed that for both genders, the frontotemporal region presents a slightly higher ageing impact than the other regions. In general, men exhibit a more pronounced effect than women. These results are consistent with previous studies based on MRI, indicating that the more pronounced ageing effects are observed in the healthy brain for men than for women. Remarkably, our approach can be used for broader cross-modal applications, including MEG data. This study advances M/EEG-based age prediction and underscores the versatility and efficacy of our proposed method.