FAPEX: Fractional Amplitude-Phase Expressor for Robust Cross-Subject Seizure Prediction¶
Conference: NEURIPS2025
arXiv: 2511.03263
Code: To be confirmed
Area: Medical Imaging
Keywords: seizure prediction, EEG, fractional Fourier transform, state-space model, phase-amplitude coupling
TL;DR¶
This paper proposes FAPEX, a framework that achieves adaptive time-frequency decomposition via a learnable Fractional Neural Frame Operator (FrNFO), combined with Amplitude-Phase Cross-Encoding (APCE) and Spatial Correlation Aggregation (SCA). FAPEX comprehensively outperforms 33 baseline methods across 12 cross-species, cross-modality seizure prediction benchmarks.
Background & Motivation¶
- Epilepsy affects over 50 million people worldwide; accurate seizure prediction is critical for timely clinical intervention.
- Most existing methods are subject-specific, requiring large amounts of labeled data for each new patient and failing to generalize across patients, which severely hinders large-scale clinical deployment.
- Subject-Agnostic Seizure Prediction (SASP) is a clinically more valuable setting, yet it faces three core challenges:
- Conventional CNN/Transformer architectures exhibit spectral bias, favoring low-frequency components and struggling to capture high-frequency oscillations (HFOs) as key biomarkers.
- Seizures involve abnormal phase-amplitude coupling (PAC), but existing models typically process time-domain and frequency-domain amplitude information separately, neglecting phase–amplitude interactions.
- Substantial inter-patient variability in electrode placement, implantation strategies, and brain region coverage leads to severe channel heterogeneity.
Core Problem¶
How to design a unified, subject-agnostic seizure prediction model that generalizes robustly across species (human/rat/canine/macaque) and acquisition modalities (Scalp-EEG/SEEG/ECoG/LFP), while effectively capturing both high- and low-frequency biomarkers as well as phase-amplitude coupling relationships?
Method¶
Overall Architecture¶
FAPEX consists of three core modules: FrNFO backbone encoder → Amplitude-Phase Cross-Encoding (APCE) → Spatial Correlation Aggregation (SCA).
1. Input Patchification¶
- Multi-channel neural signals \(\boldsymbol{X} \in \mathbb{R}^{C \times T}\) are segmented into non-overlapping patches of fixed length \(\tau\).
- Each patch is projected into a \(d_{\text{model}}\)-dimensional feature space via a shared linear embedding.
- This design renders the model independent of electrode count and spatial arrangement.
2. Fractional Neural Frame Operator (FrNFO)¶
- Core innovation: Combines the Fractional Fourier Transform (FrFT) with learnable Weyl-Heisenberg frames.
- FrFT continuously interpolates between the time domain (\(\theta=0\)) and frequency domain (\(\theta=\pi/2\)) via the fractional parameter \(\theta\), but conventional FrFT is constrained by fixed chirp functions and is sensitive to deformations.
- FrNFO addresses these limitations by:
- Generating adaptive window functions via an implicit MLP parameterized with Hermite polynomials and sinusoidal activations.
- Introducing learnable fractional orders \(\boldsymbol{\theta} \in (0, \pi)^{d_{\text{model}}}\), where each feature channel independently controls time-frequency resolution.
- Performing filtering in the fractional domain via the fractional convolution theorem: \(\hat{\boldsymbol{X}}_{:,k} = \exp(-\pi i \omega^2 \cot\theta_k) \odot \mathcal{F}_{\theta_k}(\boldsymbol{X}_{:,k}) \odot \mathcal{F}_{\theta_k}(\boldsymbol{\Psi}_{:,k})\)
- The output naturally decomposes into amplitude and phase components in complex-valued representation.
- Theoretical guarantee: From a scattering transform perspective, the amplitude representation of FrNFO is shown to possess provable robustness.
3. Amplitude-Phase Cross-Encoding (APCE)¶
- A bidirectional state-space model (Bidirectional SSM) is employed to construct a cross-attention mechanism.
- Phase BSSM: Takes phase embeddings as input while amplitude embeddings provide state-space parameters (\(\boldsymbol{B}, \boldsymbol{C}\)), capturing the modulation of amplitude by phase.
- Amplitude BSSM: Roles are reversed — amplitude serves as the query while phase provides context.
- Features are fused via residual connections, yielding representations that encode phase-amplitude coupling information.
4. Spatial Correlation Aggregation (SCA)¶
- Linear attention is used to model global cross-electrode dependencies with \(O(C)\) complexity (linear in the number of channels).
- The feature map \(\phi\) is implemented as a single-layer MLP: \(\phi_{\text{MLP}}(\boldsymbol{x}) = \exp(\boldsymbol{W}_1^\top \boldsymbol{x})\)
- Local spatiotemporal patterns are aggregated via a \(3 \times 3\) depthwise convolution gating mechanism.
Key Experimental Results¶
Experimental Scale¶
- 12 benchmark datasets: Covering 4 species (human, rat, canine, macaque) and 4 acquisition modalities (Scalp-EEG, SEEG, ECoG, LFP).
- 33 baseline methods: 23 supervised + 10 self-supervised.
- Evaluation protocol: Subject-Agnostic Nested Cross-Validation (SANCV).
Core Results (Supervised FAPEX-Base)¶
| Dataset | SEN | F1 | AUROC |
|---|---|---|---|
| Beirut (Scalp-EEG, Human) | 84.7 | 84.3 | 85.8 |
| Canine (ECoG, Canine) | 86.0 | 84.7 | 74.5 |
| FMCE (ECoG/SEEG, Human) | 88.8 | 90.7 | 97.2 |
| cTLE-RatLFP (LFP, Rat) | 81.8 | 83.2 | 91.2 |
| KAIME (EEG+SEEG, Macaque) | 87.0 | 95.6 | 90.1 |
| PCS (Scalp-EEG, Human) | 91.5 | 91.5 | 96.3 |
Self-Supervised Pretraining Gains (FAPEX-Base SSL)¶
- After pretraining, F1 improves by a further 2–10 percentage points on most datasets.
- On the IESS dataset, F1 increases from 72.4 → 84.9, a gain of 12.5 percentage points.
- On the PCS dataset, F1 improves from 91.5 → 95.0.
Cross-Domain Transfer¶
- In source-only transfer settings, FAPEX achieves relative F1 improvements typically exceeding 30% over Neuro-BERT and CBraMod.
- Even under target-domain-labeled CDAC/MME protocols, FAPEX matches or surpasses most baselines across scenarios.
Highlights & Insights¶
- Theoretical rigor: FrNFO is not merely an engineering innovation but possesses provable robustness from a scattering transform perspective.
- Exceptionally comprehensive experiments: 12 datasets, 4 species, 33 baselines, and multiple transfer learning settings constitute one of the largest comparative studies in the seizure prediction field.
- Interpretability: Visualization of FrNFO frequency responses shows progressive sub-band refinement with increasing depth, with both high- and low-frequency components effectively amplified, overcoming the low-frequency bias of conventional models.
- Compact architecture: Rather than scaling model size, high performance is achieved through mathematically principled design.
Limitations & Future Work¶
- Channel alignment preprocessing maps all data to 64 channels; this hard constraint may discard original channel layout information.
- Several proprietary datasets (AGS, ATLE, IESS, etc.) are not publicly reproducible.
- The paper does not report detailed comparisons of inference latency or parameter counts, and discussion of feasibility for wearable device deployment is insufficient.
- Ablation studies are only briefly mentioned in the appendix; the independent contributions of core modules require more transparent quantification.
Related Work & Insights¶
- vs. Medformer / TSLANet (multi-scale temporal models): FAPEX exceeds F1 by 5–10 percentage points on the vast majority of datasets, with particularly notable advantages in cross-species scenarios.
- vs. EEG foundation models (Neuro-BERT / CBraMod / LaBraM): FAPEX substantially outperforms these models even with identical pretraining data, indicating that gains stem from architectural design rather than pretraining scale.
- vs. SeizureFormer (epilepsy-specific model): FAPEX achieves 10–20 percentage point advantages in both SEN and F1.
- vs. frequency-domain methods (FreTS / NFM / ATFNet): The fractional-order design of FrNFO significantly outperforms fixed-frequency basis transforms on non-stationary signals.
Inspiration & Connections¶
- The learnable fractional-order time-frequency decomposition underlying FrNFO can be extended to other biosignal analysis tasks such as ECG and EMG.
- The SSM-based APCE architecture provides a paradigm for other signal processing tasks requiring joint modeling of amplitude and phase.
- The success of cross-species generalization implies that pre-ictal neural dynamics may share universal patterns across species.
Rating¶
- Novelty: 9/10 — An innovative combination of fractional frame theory and deep learning, with solid theoretical foundations.
- Experimental Thoroughness: 10/10 — 12 datasets, 4 species, 33 baselines, and multiple transfer settings; exceptionally comprehensive.
- Writing Quality: 8/10 — Generally clear, but notation is mathematically dense with occasional inconsistencies.
- Value: 9/10 — A significant contribution to the seizure prediction field with high potential for clinical translation.