Assessing the Robustness of Deep Learning Based Brain Age Prediction Models Across Multiple EEG Datasets

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Large electroencephalography (EEG) datasets are increasingly available, offering potential for deep learning (DL) in clinical applications.
• Dataset shifts, caused by population and hardware variations, significantly degrade DL model performance in decoding cognitive and pathological states.
• Investigating the generalization of DL models across diverse EEG datasets is crucial for reliable clinical translation.

Purpose of the Study:

• To systematically evaluate the generalization of deep learning models for age decoding using EEG data across different datasets.
• To identify key hyperparameters and pre-processing strategies that enhance model robustness against dataset shifts.
• To establish a benchmark for future research on improving the generalization of EEG-based DL models.

Main Methods:

• Utilized five distinct EEG datasets with two cross-validation strategies: leave-one-dataset-out (LODO) and leave-one-dataset-in (LODI).
• Tested 1805 hyperparameter configurations, exploring variations in DL architectures and data pre-processing techniques.
• Assessed model performance using Pearson's r and R-squared metrics, with additional analysis on adjusting model intercepts.

Main Results:

• Deep learning models demonstrated the ability to learn generalizable age-related EEG patterns, with performance varying by dataset pair.
• The frequency range of 1-45Hz was identified as the most critical hyperparameter for generalization, outperforming single frequency bands.
• Adjusting model intercepts with target dataset average age improved R-squared scores in specific scenarios, highlighting the impact of dataset characteristics.

Conclusions:

• Deep learning models can generalize age-related EEG patterns across diverse datasets, though dataset shifts present a significant challenge.
• Hyperparameter tuning, particularly the use of a broad frequency range (1-45Hz), is essential for robust EEG-based age decoding.
• Findings provide a benchmark for developing more resilient DL models for clinical applications using heterogeneous EEG data.