PHINN-EEG: Topological Time-Series Analysis of Dream-State EEG -- Dynamic Betti Curves for Dream Content Classification and Topology-Conditioned Neural Signal Synthesis

Technical Summary: PHINN-EEG

Problem Statement
Current computational approaches to detecting dream mentation from electroencephalography (EEG) rely primarily on Power Spectral Density (PSD) and statistical moment features (e.g., catch22). While state-of-the-art methods achieve an Area Under the Receiver Operating Characteristic Curve (AUC) of approximately 0.70 on the DREAM database, they treat dreaming as a problem of spectral energy. The authors argue this approach has a fundamental limitation: it measures how much brain activity is present but fails to capture the geometric shape of that activity. The paper posits that dream states involve qualitatively different network integration architectures (predicted by Integrated Information Theory and Global Workspace Theory) that are invisible to spectral decomposition but measurable via topological invariants.

Methodology: The PHINN-EEG Pipeline
The paper introduces PHINN-EEG (Persistent Homology Inspired Neural Network for EEG), a framework that shifts from spectral energy analysis to phase-space geometry. The pipeline consists of four main stages:

1. Delay Embedding & Joint Point Cloud Construction:

   • Multi-channel EEG epochs (30 seconds pre-awakening) are processed using Takens delay embeddings.
   • A sliding-window approach is used with a 5-second window (500 samples at 100 Hz) and 80% overlap.
   • Per-channel delay vectors are concatenated to form a joint point cloud in $\mathbb{R}^{K \times d}$, where $K$ is the channel count (6–18) and $d$ is the embedding dimension (default $d=7$, with sensitivity sweeps up to $d=15$).
   • Note: The authors explicitly acknowledge that this concatenation is a heuristic; it does not guarantee a diffeomorphic reconstruction of the underlying neural attractor but captures cross-channel phase relationships.

2. Topological Feature Extraction (Dynamic Betti Curves):

   • Vietoris–Rips Filtration is applied to the point clouds. To ensure computational tractability and avoid intractable simplex counts, the filtration is bounded:
     • The primary Betti curve threshold $\epsilon^*(t)$ is set at the 10th percentile of pairwise distances within each sub-window.
     • Persistence landscapes and entropy are computed with a cap at the 25th percentile.
   • The system extracts Dynamic Betti Curves ($\beta_0(t), \beta_1(t), \beta_2(t)$) representing connected components, loops, and voids, respectively.
   • Additional features include Persistence Landscapes, Persistent Entropy, the Euler Characteristic trajectory, and a Certified Persistence Ratio (CPR) for noise robustness.
   • Total feature dimensionality is approximately 340 real-valued features per epoch.

3. Classification Architecture:

   • Model 1 (Topology Ensemble): An XGBoost classifier trained on the full topological feature vector.
   • Model 2 (Raw EEG Ensemble): An XGBoost classifier trained on standard PSD and catch22 features (approx. 220 features) to serve as a baseline.
   • Model 2b (Dimensionality-Matched Baseline): Model 2 expanded with additional spectral bins to match Model 1's feature count, ensuring comparisons are not biased by feature-space capacity.
   • Model 3 (Temporal Fusion): A CNN-BiLSTM architecture using cross-attention to fuse raw EEG and topological features.
   • Validation: The study employs a Leave-One-Dataset-Out (LODO) cross-validation scheme across six open-access subsets of the DREAM database. Crucially, primary hypothesis testing is restricted to "harmonised-reference" folds (where training and test sets share the same reference type, e.g., unipolar-to-Cz vs. bipolar) to avoid confounding signal representation with dream content.

4. Generative Synthesis:

   • A Topology-Conditioned Rectified Flow model is introduced to synthesize dream-state EEG.
   • The model is conditioned on the topological feature vector ($\beta^*$) via a cross-attention mechanism within a U-Net backbone.
   • Training uses Rectified Flow Matching (LRM) loss. A topological fidelity metric ($L_{topo}$) is used for offline validation but is not a differentiable training loss due to the non-differentiability of Vietoris–Rips filtration.
   • Baselines include unconditional flow, spectral-conditioned flow (dimensionality-matched), EEG-GAN, VAE, and a latent diffusion model trained from scratch.

Key Contributions

1. First Topological Framework for Dream Detection: Introduces the first application of persistent homology to dream content classification on the DREAM database.
2. Dynamic Betti Curves: Formalizes the extraction of $\beta_0, \beta_1, \beta_2$ from multi-channel Takens embeddings as discriminative features for dream vs. dreamless states.
3. Topology-Conditioned Synthesis: Proposes a rectified flow model that generates EEG conditioned on topological invariants, evaluated against spectral and unconditional baselines.
4. Exploratory Hypothesis Space: Presents candidate "Betti transition archetypes" linking topological patterns (e.g., sustained $\beta_1$ loops) to phenomenological dream categories (e.g., structured narratives vs. fragmented recall), explicitly framed as unvalidated hypotheses pending empirical confirmation.

Projected Results and Performance Targets
The paper explicitly labels all performance figures as projections based on verified benchmarks from related TDA-EEG tasks (sleep staging and meditative EEG), not empirical results from the DREAM database yet.

• REM Dream Detection: Targets an AUC of 0.82–0.90, compared to the current benchmark of 0.700 (Wong et al.).
• NREM Dream Detection: Targets an AUC of 0.72–0.78, compared to the baseline of 0.586.
• Generative Quality: Projects a Scenario Coverage Rate (SCR) of 60–70% (vs. 3–5% random) and lower Fréchet EEG Distance (FED) than existing baselines.
• Ablation: The authors anticipate that a PCA-reduced topological model will significantly outperform the raw EEG baseline, confirming that the gain is due to topological content rather than feature dimensionality.

Significance and Claims
The paper claims a paradigm shift in neural rare-event detection, moving from asking "how much brain activity is present?" to "what geometric shape does that activity possess?"

• Theoretical Shift: It argues that topology captures the "shape of dreaming" (persistent loops, voids, and integration) which spectral methods miss.
• Validation Status: The authors are modest about their claims. They state that the "Geometric Dream Hypothesis" and the specific mapping of Betti numbers to dream content are exploratory and require empirical validation on the DREAM database.
• Limitations: The paper acknowledges several constraints:
  • Volume Conduction: The inability to apply surface Laplacians on sparse montages (6–18 channels) means zero-lag spatial mixing remains a confound, addressed only indirectly via MIAAFT surrogates and channel perturbation controls.
  • Signal Representation: The pipeline operates on mathematically different signal types (unipolar vs. bipolar) across datasets, limiting cross-dataset generalization claims until harmonized-reference folds are used.
  • Real-Time Feasibility: While sub-window latency is low (<300 ms), full-epoch feature extraction currently takes several seconds, meaning real-time BCI deployment is not yet established.

In summary, PHINN-EEG proposes a novel topological approach to dream detection and synthesis. While the projected performance targets are ambitious (AUC 0.82–0.90), the paper frames these as working targets grounded in prior TDA success in other EEG domains, with the immediate next step being the empirical execution and validation of the pipeline on the DREAM database.