BioFormer: New AI Model Aligns Brain Signals Across Subjects
A new AI architecture called BioFormer addresses the challenge of cross-subject generalization in biomedical time-series (BTS) analysis. The key innovation is the concept of 'spectral drift'—subject-specific variability in frequency components of BTS signals. BioFormer's Frequency-Band Alignment Module (FBAM) generates band-wise modulation factors from spectral distributions to adaptively align amplitude and phase across subjects. This approach explicitly models variability rather than suppressing it implicitly, potentially improving performance on unseen subjects. The paper is available on arXiv (2605.22468).
Key facts
- BioFormer introduces spectral drift to characterize subject-specific variability in biomedical time-series.
- The Frequency-Band Alignment Module (FBAM) generates band-wise modulation factors.
- FBAM adaptively adjusts amplitude and phase to align spectral structure.
- Cross-subject generalization means training on some subjects and testing on unseen ones.
- Existing methods suppress variability implicitly through model building or adversarial learning.
- BTS signals under the same label share consistent oscillatory structure.
- Subject-dependent magnitude or phase shifts occur in specific frequency components.
- The paper is published on arXiv with ID 2605.22468.
Entities
Institutions
- arXiv