Lifespan Brain Age Prediction based on Multiple EEG Oscillatory Features and Sparse Group Lasso

Shiang Hu^1,2*Xue Xiang^1,2Xiaolong Huang^1,2Yan Lu³Xulai Zhang⁴Dezhong Yao⁵Pedro Antonio Valdes-Sosa⁵

• ¹School of Computer Science and Technology, Anhui University, Hefei, Anhui Province, China
• ²Anhui Province Key Laboratory of Multimodal Cognitive Computation, Anhui University, Hefei, Anhui Province, China
• ³Stony Brook Institute at Anhui University, Hefei, Anhui Province, China
• ⁴Fourth People’s Hospital of Hefei, Anhui Mental Health Center, Hefei, Anhui Province, China
• ⁵Key Laboratory for NeuroInformation of Ministry of Education, School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, Sichuan Province, China

The neural dynamics underlying cognition and behavior change greatly during the lifespan of brain development and aging. EEG is a promising modality due to its high temporal resolution in capturing neural oscillations. Precise prediction of brain age (BA) based on EEG is crucial to screening high-risk individuals from large cohorts. However, the lifespan representation of the EEG oscillatory features (OSFs) is largely unclear, limiting practical BA applications in clinical scenarios. This study aims to build an interpretable BA prediction model through prior knowledge and sparse group lasso. Based on the multinational cross-spectral (MNCS) dataset that covers 5-97 years, (1) we extracted four groups of OSFs, such as aperiodic parameters, periodic parameters, power-ratio, and relative power; (2) the OSFs trajectories evolving with age and the OSF importance topographies were mapped using the generalized additive model for location, scale and shape (GAMLSS) and Pearson's correlation coefficient (PCC); (3) the inter-oscillatory dependency coefficients (ODCs) were extracted by the sparse group lasso; 4) the fusion of OSFs and ODCs was flattened and fed into a three-layer fully connected neural network (FCNN); the FCNN interpretability was analyzed by Layerwise Relevance Propagation and 10-fold cross-validation. The results showed that the FCNN model that incorporated ODC significantly improved the prediction of BA (MAE=2.95 years, R 2 =0.86) compared to the use of only OSF (MAE = 3.44 years, R 2 = 0.84). In general, this study proposed a BA prediction model by systematically employing OSFs and highlighting the interpretability of the model, which holds broad promise by integrating normative modeling for precise individual stratification.