Di Zhang,1,2 Yichong She,1,2 Jinbo Sun,1,2 Yapeng Cui,1,2 Xuejuan Yang,1,2 Xiao Zeng,1,2 Wei Qin1,2 

1Engineering Research Center of Molecular and Neuro Imaging of the Ministry of Education, School of Life Science and Technology, Xidian University, Xi’an, Shaanxi, 710126, People’s Republic of China; 2Intelligent Non-Invasive Neuromodulation Technology and Transformation Joint Laboratory, Xidian University, Xi’an, People’s Republic of China

Correspondence: Xiao Zeng, Xidian University, 266 Xifeng Road, Xinglong, Xi’an, Shaanxi, People’s Republic of China, 710126, Email xiaozeng1024@gmail.com

Purpose: This study aims to improve brain age estimation by developing a novel deep learning model utilizing overnight electroencephalography (EEG) data.
Methods: We address limitations in current brain age prediction methods by proposing a model trained and evaluated on multiple cohort data, covering a broad age range. The model employs a one-dimensional Swin Transformer to efficiently extract complex patterns from sleep EEG signals and a convolutional neural network with attentional mechanisms to summarize sleep structural features. A multi-flow learning-based framework attentively merges these two features, employing sleep structural information to direct and augment the EEG features. A post-prediction model is designed to integrate the age-related features throughout the night. Furthermore, we propose a DecadeCE loss function to address the problem of an uneven age distribution.
Results: We utilized 18,767 polysomnograms (PSGs) from 13,616 subjects to develop and evaluate the proposed model. The model achieves a mean absolute error (MAE) of 4.19 and a correlation of 0.97 on the mixed-cohort test set, and an MAE of 6.18 years and a correlation of 0.78 on an independent test set. Our brain age estimation work reduced the error by more than 1 year compared to other studies that also used EEG, achieving the level of neuroimaging. The estimated brain age index demonstrated longitudinal sensitivity and exhibited a significant increase of 1.27 years in individuals with psychiatric or neurological disorders relative to healthy individuals.
Conclusion: The multi-flow deep learning model proposed in this study, based on overnight EEG, represents a more accurate approach for estimating brain age. The utilization of overnight sleep EEG for the prediction of brain age is both cost-effective and adept at capturing dynamic changes. These findings demonstrate the potential of EEG in predicting brain age, presenting a noninvasive and accessible method for assessing brain aging.

Keywords: brain age, sleep polysomnography, electroencephalography, deep learning, swin transformer