Spike-Based Adaptation of EEG Foundation Models for Calibration Efficient Cross-Subject Brain-Computer Interfaces
Спайковая адаптация фундаментальных моделей ЭЭГ для эффективных по калибровке межсубъектных интерфейсов мозг–компьютер
2026-09-14
SCID: 54.1/97nry8at
Discuss with AI
EEG foundation modelscross-subject brain-computer interfacesfew-shot calibrationneuromorphic computingspike-based adaptation
Figures from the paper
Abstract (AI)
Abstract Non-invasive electroencephalography (EEG)-based brain–computer interfaces (BCIs) hold potential for assistive communication, neurorehabilitation, and human–machine interaction. However, cross-subject variability hinders deployment because most systems require subject-specific calibration or costly retraining for unseen individuals. EEG foundation models improve transferability, but downstream adaptation often remains computationally intensive for wearable BCIs. We propose a spike-based adaptation framework integrating a pretrained EEG foundation model with a lightweight neuromorphic module for cross-subject personalisation. The foundation model extracts transferable latent representations, while the event-driven spiking module captures subject-specific temporal variations through sparse computation. The framework uses limited calibration trials without modifying the pretrained backbone. It is evaluated on four EEG datasets comprising 175 participants across motor imagery, P300/ERP, SSVEP, and multi-session visual paradigms under leave-one-subject-out, cross-session, and few-shot protocols. Performance is compared with full fine-tuning, linear probing, transfer-learning approaches, lightweight adapters, and neuromorphic models. It achieved 86.7% average cross-subject accuracy and macro-F1 of 0.861, while retaining 84.9% accuracy with 10% calibration data. Compared with full fine-tuning, it reduced adaptation time by 87.6% and trainable parameters by 98.9%, required 0.42 mJ per trial for the adaptation pathway, and limited cross-session degradation to 4.3 percentage points. Subject-level improvements remained significant after Holm–Bonferroni correction (p < 0.001). These findings support calibration-efficient, resource-aware cross-subject BCI systems by effectively bridging EEG foundation models with sparse neuromorphic processing.
Key Findings
1
Across four EEG datasets involving 175 participants and multiple BCI paradigms, it achieved 86.7% average cross-subject accuracy and a macro-F1 of 0.861.
2
Compared with full fine-tuning, adaptation time decreased by 87.6% and trainable parameters by 98.9%, while the adaptation pathway consumed 0.42 mJ per trial.
3
The method limited cross-session performance degradation to 4.3 percentage points, with subject-level improvements remaining significant after Holm–Bonferroni correction (p < 0.001).
4
The proposed framework combines a pretrained EEG foundation model with a lightweight event-driven spiking module for cross-subject personalization without modifying the backbone.
5
Using only 10% of calibration data, the method retained 84.9% accuracy, demonstrating calibration-efficient adaptation for unseen subjects.
Research Object
EEG-based brain–computer interfaces for cross-subject decoding across motor imagery, P300/ERP, SSVEP, and multi-session visual paradigms
Research Subject
Calibration-efficient cross-subject personalisation, including decoding performance, adaptation efficiency, computational resource use, and cross-session robustness
Publication Details
Publication Date
2026-09-14
Journal
Publisher
ISSN
Cited by
0
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest
References available in scid.ai5
EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces2018
Deep learning with convolutional neural networks for EEG decoding and visualization2017
A review of classification algorithms for EEG-based brain–computer interfaces: a 10 year update2018
Non-Invasive Brain-Computer Interfaces: State of the Art and Trends2024
Deep Learning Approaches for EEG-Motor Imagery-Based BCIs: Current Models, Generalization Challenges, and Emerging Trends2025