| name | phinn-eeg-topological-dream-analysis |
| category | neuroscience |
| description | Topological time-series analysis methodology for dream-state EEG using Dynamic Betti curves, persistent homology, and topology-conditioned neural signal synthesis. arXiv:2607.09662 |
| created | 2026-07-13T00:00:00.000Z |
| source | arXiv:2607.09662v1 (Takahashi, Yusuf, Bhaduri, 2026-07-10) |
PHINN-EEG: Topological Dream-State EEG Analysis
Overview
PHINN-EEG (Persistent Homology Inspired Neural Network for EEG) is the first topological time-series framework for dream mentation analysis. It shifts dream detection from spectral energy (PSD) to phase-space geometry using persistent homology.
Core Methodology
1. Dynamic Betti Curves
- Sliding-window Takens delay embeddings on multichannel pre-awakening EEG epochs
- Vietoris-Rips filtrations extract topological invariants (Betti numbers β₀, β₁, β₂...)
- Dynamic Betti Curves characterize the geometric architecture of neural activity, not just energy
- Targets AUC = 0.82-0.90 vs. ~0.70 SOTA (PSD + catch22 benchmarks)
2. Topology-Conditioned Flow Matching
- Topology-conditioned rectified flow model for dream-state EEG synthesis
- Spectral-conditioned flow model used as ablation baseline to isolate topological conditioning value
- Candidate Betti transition archetypes link topology to phenomenological dream report categories
3. Dataset
- DREAM database: 1,462-awakening open-access subset (from 3,191 total awakenings, 263 participants, 20 labs)
Key Innovation
Paradigm shift: From spectral energy analysis to phase-space geometry for neural rare-event detection. The Betti curves capture topological features (connected components, loops, voids) in the reconstructed state space of EEG signals.
Implementation Steps
- Preprocessing: Extract pre-awakening EEG epochs from polysomnography data
- Takens Embedding: Apply delay embedding to reconstruct phase space (choose embedding dimension d and delay τ)
- Vietoris-Rips Filtration: Compute persistent homology across filtration parameter ε
- Betti Curve Extraction: Track β₀(t), β₁(t), β₂(t) over sliding windows
- Classification: Feed Dynamic Betti Curves (or topology-conditioned features) to a downstream classifier; use topology-conditioned rectified flow for synthesis/ablation
Activation / Triggers
phinn-eeg, betti curves, persistent homology, topological eeg, dream detection, takens embedding, vietoris-rips, topology-conditioned flow
Verification
- Outperforms PSD and catch22 baselines on the DREAM open-access subset (target AUC 0.82-0.90)
- Spectral-conditioned ablation confirms added value of topological conditioning
- Betti transition archetypes reproducible across the 20 independent laboratories subset
