Practical Guidelines for Improving Accuracy in EEG-derived Brain Age Prediction using Deep Learning
Poster No:
1318
Submission Type:
Abstract Submission
Authors:
Javier Urriola1, Leo Lebrat2,3, Ian Zajac4,5, David White6, DanaKai Bradford1, Naomi Kakoschke4
Institutions:
1Australian E-Health Research Centre, Health & Biosecurity, CSIRO, Brisbane, Australia, 2Data61, CSIRO, Brisbane, Australia, 3School of Electrical Engineering & Robotics, Queensland University of Technology, Brisbane, Australia, 4Public Health & Wellbeing, Health & Biosecurity, CSIRO, Adelaide, Australia, 5Department of Psychology, College of Education, Psychology and Social Work, Flinders University, Adelaide, Australia, 6Centre for Mental Health and Brain Sciences, School of Health Sciences, Swinburne University of Tech, Melbourne, Australia
First Author:
Javier Urriola, Dr
Australian E-Health Research Centre, Health & Biosecurity, CSIRO
Brisbane, Australia
Australian E-Health Research Centre, Health & Biosecurity, CSIRO
Brisbane, Australia
Co-Author(s):
Leo Lebrat, Dr
Data61, CSIRO|School of Electrical Engineering & Robotics, Queensland University of Technology
Brisbane, Australia|Brisbane, Australia
Data61, CSIRO|School of Electrical Engineering & Robotics, Queensland University of Technology
Brisbane, Australia|Brisbane, Australia
Ian Zajac, Dr
Public Health & Wellbeing, Health & Biosecurity, CSIRO|Department of Psychology, College of Education, Psychology and Social Work, Flinders University
Adelaide, Australia|Adelaide, Australia
Public Health & Wellbeing, Health & Biosecurity, CSIRO|Department of Psychology, College of Education, Psychology and Social Work, Flinders University
Adelaide, Australia|Adelaide, Australia
David White, A/Prof
Centre for Mental Health and Brain Sciences, School of Health Sciences, Swinburne University of Tech
Melbourne, Australia
Centre for Mental Health and Brain Sciences, School of Health Sciences, Swinburne University of Tech
Melbourne, Australia
DanaKai Bradford
Australian E-Health Research Centre, Health & Biosecurity, CSIRO
Brisbane, Australia
Australian E-Health Research Centre, Health & Biosecurity, CSIRO
Brisbane, Australia
Introduction:
Electroencephalography (EEG)-derived brain age predictions can identify individuals at risk for age-related neurological disorders (Cole & Franke, 2017). Machine learning models show potential for achieving accurate predictions (Al Zoubi et al., 2018; Kounios et al., 2024), with reusable benchmarks (Engemann et al., 2022), highlighting the value of standardised pipelines across populations. Despite these advancements, optimal electrode density, recording duration, and the relative contribution of specific brain regions remain insufficiently explored. To address these gaps, we systematically compared Deep4Net (deep) and ShallowFBCSPNet (shallow) learning methods on open-source, demographically diverse EEG datasets. Our study aimed to optimise key EEG acquisition parameters and compare deep learning architectures for accurate brain age prediction.
Methods:
Single-session resting-state EEG data were obtained from three datasets (n=1836): LEMON (n=211; 61 channels; Mean age 39.2 years, range 22.5–77.5), the Temple University Abnormal EEG Corpus (TUAB; n=1017; 21 channels; Mean age 43.9 years, range 1–95), and the Dortmund Vital Study (DVS; n=608; 64 channels; Mean age 44 years, range 20–70). Data were pre-processed using MNE-Python following published reusable benchmark steps, including bandpass filtering (0.1–49 Hz), segmentation into 10-second epochs and automated artefact rejection (Jas et al., 2017). Deep and shallow convolutional neural networks (Schirrmeister et al., 2017) were trained to predict chronological age from pre-processed EEG recordings. Our analysis focused on age prediction accuracy, varying three key factors: electrode density (7–64 channels) categorised as low (<19 electrodes), clinical standard (19–25 electrodes), and high density (>32 electrodes); recording duration (1–15 mins); and contributions from six anatomical regions. Prediction accuracy was evaluated using mean absolute error (MAE) for average deviation in years, coefficient of determination (R²) for predictive power, bias for systematic over/underestimation, and limits of agreement (LoA) to capture 95% of prediction errors.
Results:
Electrode density analysis revealed diminishing returns beyond clinical-standard montages (19-25 electrodes) for both shallow and deep architecture, with MAE improvements <0.5 years for higher densities [Fig. 1a]. Regional analyses demonstrated strongest predictive power in centro-temporal (Deep: R²=0.52; Shallow: R²=0.58) and temporal regions across both architectures [Fig. 1b]. Recording duration analysis showed optimal performance for the shallow architecture at 3-5 mins (MAE=8.3–8.4 years, R²=0.63–0.64) with minimal bias (-0.33 – -0.16 years) and no substantial improvements with longer recordings [Fig. 2a]. The deep architecture showed continued benefits from longer recordings, with best performance at 15 mins (MAE=7.6 years, R²=0.66), although most improvements were achieved by 10 mins with only marginal gains (+0.4 years MAE reduction) beyond this duration [Fig. 2a,b]. Notably, shallow models achieved near-zero bias at shorter durations, while deep models retained a persistent positive bias at longer durations.
·Figure 1. Impact of electrode density and regional sources on EEG-based brain age prediction. (a) Mean Absolute Error (MAE) in years as a function of electrode number for deep (left) and shallow (righ
·Figure 2. Impact of recording duration on brain age prediction performance. (a) Mean Absolute Error (MAE) in years as a function of recording duration for deep (left) and shallow (right) architectures
Conclusions:
Our findings offer practical guidelines for EEG-derived brain age prediction. A standard clinical montage (21-25 electrodes) suffices for both architectures, with deep showing minimal gains from higher densities. Strong predictive power in centro-temporal and temporal regions aligns with multimodal imaging findings (Guan et al., 2024) , validating their biological relevance in aging-related EEG patterns. Recording duration analysis revealed trade-offs: shallow models achieve stable, nearly unbiased predictions with shorter recordings (3-5 mins), while deep models provide higher accuracy (MAE=7.6 vs 8.1 years) but require longer recordings (≥10 mins) and show systematic overestimation. These trade-offs highlight the need to balance accuracy, efficiency, and bias when selecting models for clinical applications.
Lifespan Development:
Aging 2
Modeling and Analysis Methods:
Classification and Predictive Modeling
EEG/MEG Modeling and Analysis 1
Methods Development
Novel Imaging Acquisition Methods:
EEG
Keywords:
Aging
Electroencephaolography (EEG)
Machine Learning
Other - Brain age prediction
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
Which processing packages did you use for your study?
Provide references using APA citation style.
Cole, J. H., & Franke, K. (2017). Predicting age using neuroimaging: innovative brain ageing biomarkers. Trends in neurosciences, 40(12), 681-690.
Engemann, D. A., Mellot, A., Höchenberger, R., Banville, H., Sabbagh, D., Gemein, L., Ball, T., & Gramfort, A. (2022). A reusable benchmark of brain-age prediction from M/EEG resting-state signals. Neuroimage, 262, 119521.
Guan, S., Jiang, R., Meng, C., & Biswal, B. (2024). Brain age prediction across the human lifespan using multimodal MRI data. GeroScience, 46(1), 1-20.
Jas, M., Engemann, D. A., Bekhti, Y., Raimondo, F., & Gramfort, A. (2017). Autoreject: Automated artifact rejection for MEG and EEG data. Neuroimage, 159, 417-429. https://doi.org/10.1016/j.neuroimage.2017.06.030
Kounios, J., Fleck, J. I., Zhang, F., & Oh, Y. (2024). Brain-age estimation with a low-cost EEG-headset: effectiveness and implications for large-scale screening and brain optimization. Front Neuroergon, 5, 1340732.
Schirrmeister, R. T., Springenberg, J. T., Fiederer, L. D. J., Glasstetter, M., Eggensperger, K., Tangermann, M., Hutter, F., Burgard, W., & Ball, T. (2017). Deep learning with convolutional neural networks for EEG decoding and visualization. Human brain mapping, 38(11), 5391-5420.