{
  "abstract": "EEG foundation models, pre-trained on large-scale unlabelled EEG data, have emerged as a promising direction towards learning generalizable EEG representations. Despite showing positive results in data-rich regimes, they often fail to outperform significantly smaller supervised models in low-resource settings compared to fully supervised models. We provide a mechanistic account of this shortcoming, attributing it to a fundamental mismatch between reconstruction-based pretext tasks and the idiosyncratic spectral structure of EEG signals, which decompose into distinct high-power aperiodic and low-power oscillatory components. Using controlled, synthetically-generated EEG inputs, we demonstrate that EEG foundation model embeddings are biased to capture the aperiodic components of the EEG signal while under-representing oscillatory components, particularly at higher frequencies. Additionally, linear probe evaluations on real-world BCI datasets further reveal that embeddings encode subject identity more strongly than task-relevant information, thereby reinforcing the low-frequency and aperiodic component bias in foundation model embeddings trained primarily on reconstruction based objectives. Together, these findings elucidate a failure mode in reconstruction based EEG foundation models and motivate future work to incorporate auxiliary losses explicitly targeting high-frequency oscillatory structure as a path toward more capable and generalizable EEG representations.",
  "abstract_shown": true,
  "affiliations": [
    "Department of Electrical and Computer Engineering University of Southern California Thomas Lord Department of Computer Science",
    "University of Southern California Department of Mathematical Sciences",
    "Aalborg University"
  ],
  "article_type": null,
  "arxiv": null,
  "author_total": 10,
  "authors": [
    {
      "affs": [
        0
      ],
      "name": "Aditya Kommineni",
      "orcid": "0009-0002-0056-8336"
    },
    {
      "affs": [
        1
      ],
      "name": "Emily Zhou",
      "orcid": "0009-0009-5004-6588"
    },
    {
      "affs": [
        1
      ],
      "name": "Kleanthis Avramidis",
      "orcid": "0000-0003-0308-795X"
    },
    {
      "affs": [
        2
      ],
      "name": "Simon Bock Segaard"
    },
    {
      "affs": [
        2
      ],
      "name": "Jeppe Roden Münster"
    },
    {
      "name": "Andreas Peter Juhl Hansen"
    },
    {
      "affs": [
        0
      ],
      "name": "Takfarinas Medani",
      "orcid": "0000-0003-0774-067X"
    },
    {
      "affs": [
        0
      ],
      "name": "Tiantian Feng",
      "orcid": "0000-0002-2053-9068"
    },
    {
      "affs": [
        0
      ],
      "name": "Richard M. Leahy",
      "orcid": "0000-0002-7278-5471"
    },
    {
      "affs": [
        0,
        1
      ],
      "name": "Shrikanth Narayanan"
    }
  ],
  "bibtex": "https://research.draconitic.com/paper/a5dhvc4hsqdzv.bib",
  "cited_by": 5,
  "doi": "10.48550/arxiv.2605.26434",
  "fulltext_url": null,
  "id": "a5dhvc4hsqdzv",
  "issns": [
    "2331-8422"
  ],
  "landing_url": "https://arxiv.org/abs/2605.26434",
  "language": "en",
  "licence": "CC-BY-NC-SA",
  "licence_label": "CC BY-NC-SA",
  "licence_source": "journal",
  "licence_url": "https://creativecommons.org/licenses/by-nc-sa/4.0/",
  "metadata_licence": null,
  "native_subjects": [
    [
      "openalex",
      "2805|EEG and Brain-Computer Interfaces"
    ]
  ],
  "openalex": "W7162534101",
  "pdf_url": "https://arxiv.org/pdf/2605.26434",
  "pmcid": null,
  "pmid": null,
  "published": "2026-05-26",
  "publisher": "Cornell University",
  "ris": "https://research.draconitic.com/paper/a5dhvc4hsqdzv.ris",
  "sources": [
    {
      "first_seen": "2026-11-10T13:21:11Z",
      "native_id": "W7162534101",
      "source": "openalex",
      "url": "https://arxiv.org/abs/2605.26434"
    }
  ],
  "subjects": [
    {
      "field": "biology",
      "origin": "openalex:2805|EEG and Brain-Computer Interfaces",
      "subfield": "neuroscience"
    }
  ],
  "title": "Aperiodic and Low-Frequency Spectral Bias in Reconstruction based EEG Foundation Models",
  "title_text": "Aperiodic and Low-Frequency Spectral Bias in Reconstruction based EEG Foundation Models",
  "type": "preprint",
  "updated_at": "2026-11-10T13:21:11Z",
  "url": "https://research.draconitic.com/paper/a5dhvc4hsqdzv",
  "venue": "arXiv (Cornell University)",
  "versions": []
}