Aperiodic and Low-Frequency Spectral Bias in Reconstruction based EEG Foundation Models

Researchers identify a fundamental weakness in EEG foundation models: reconstruction-based pretraining causes these models to heavily bias toward aperiodic signal components while neglecting high-frequency oscillatory patterns critical for brain-computer interfaces. This spectral mismatch explains why large pretrained models underperform smaller supervised alternatives in low-resource settings.

EEG foundation models represent an emerging approach to leverage massive unlabeled neuroimaging data for building generalizable brain signal representations. However, this study reveals a critical mechanistic failure: the standard reconstruction objective—predicting missing or corrupted EEG segments—inadvertently amplifies bias toward low-frequency, high-power aperiodic components while suppressing oscillatory information essential for downstream tasks. The researchers validated this through synthetic experiments demonstrating clear spectral bias and real-world BCI dataset analysis showing embeddings encode subject identity rather than task-relevant information.

This finding matters because foundation models have generated significant excitement in machine learning, with the assumption that massive pretraining transfers effectively across domains. The EEG domain presents unique challenges due to its dual-component spectral structure—strong aperiodic backgrounds can dominate learning signals if objectives aren't carefully designed. The discovery that reconstruction-based approaches inherently privilege these components suggests fundamental incompatibilities between common pretraining paradigms and neurophysiological signal structure.

For the neurotechnology and brain-computer interface industry, this work identifies a critical optimization opportunity. Current approaches may be leaving substantial performance on the table by using reconstruction objectives that mathematically favor the wrong signal components. Developers building EEG analysis systems should evaluate whether auxiliary losses targeting oscillatory content could improve generalization, particularly for low-data scenarios where foundation model advantages theoretically shine brightest.

The path forward involves designing hybrid objectives that balance reconstruction with explicit oscillatory component preservation. This mechanistic understanding of failure modes represents important foundational work that could reshape how neuroimaging foundation models are trained and evaluated.

• →Reconstruction-based EEG foundation models exhibit systematic bias toward aperiodic components while underrepresenting high-frequency oscillations critical for brain-computer interfaces.
• →Large pretrained EEG models encode subject identity more strongly than task-relevant information, limiting their effectiveness in low-resource settings compared to supervised baselines.
• →The mismatch between reconstruction objectives and EEG spectral structure explains why foundation models fail to deliver expected transfer learning benefits in neuroscience applications.
• →Synthetic and real-world BCI experiments confirm that spectral bias is a fundamental mechanistic failure rather than a data or architecture issue.
• →Future EEG foundation models require auxiliary losses explicitly targeting oscillatory structure to improve generalization and match or exceed supervised model performance.