Research map / Label-efficient and robust learning / Self-supervised and contrastive
FAPEX: Fractional Amplitude-Phase Expressor for Robust Cross-Subject Seizure Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FAPEX, a seizure prediction architecture with a learnable fractional neural frame operator (FrNFO) for time-frequency decomposition, amplitude-phase cross-encoding via bidirectional Mamba-style SSMs, and linear-attention spatial aggregation, evaluated on 12 cross-species EEG/ECoG/SEEG/LFP benchmarks.
Problem
Subject-agnostic seizure prediction from electrophysiological signals is hard due to inter-subject spectral variability, low-frequency spectral bias of CNNs/Transformers, and heterogeneous electrode layouts.
Approach
Combines a learnable fractional-order Weyl-Heisenberg frame filterbank with implicit-MLP-generated Hermite-sine windows and per-channel learnable fractional orders, then models the resulting phase and amplitude streams with cross-directional state-space blocks and channel-wise linear attention.
Key results
- FAPEX-Base (SSL) achieves top-1 SEN and F1 on all 12 datasets under subject-dependent evaluation per the paper's claim, with the paper citing ~10% F1/sensitivity gains over baselines.
- FAPEX-Base (SL) reaches median F1 of 84.3 on Beirut and 84.7 on Canine, outperforming ModernTCN and other supervised baselines in the tables.
- In out-of-domain source-only transfer, FAPEX-Base reports relative F1 gains often exceeding 30% over Neuro-BERT and CBraMod.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FAPEX-Base instead of the strongest listed supervised baseline | supervised SASP · Beirut | F1: 84.1 vs 84.3 | Helped |
| Use FAPEX-Base instead of the strongest listed supervised baseline | supervised SASP · Canine | F1: 84.7 vs 84.7 | No clear effect |
| Use FAPEX-Base instead of the strongest listed self-supervised baseline | self-supervised SASP · Beirut | F1: 86.0 vs 87.3 | Helped |
| Use FAPEX-Base instead of the strongest listed self-supervised baseline | self-supervised SASP · Canine | F1: 94.5 vs 94.9 | Helped |
| Add self-supervised pretraining and finetuning | FAPEX-Base · Beirut | F1: 84.3 -> 87.3 | Helped |
| Add self-supervised pretraining and finetuning | FAPEX-Base · ATLE | F1: 98.8 -> 98.0 | Hurt |
| Adding FrNFO, APCE, and SCA together in FAPEX | FAPEX · 12 seizure prediction datasets | F1 | Helped |
Limitations
- In-house datasets are available only on request.
- Code is promised upon publication but not released at submission.
- Channel harmonization to 64 channels via rejection and duplication is applied during preprocessing.
Experimental setup
Datasets: FMCE HUP RESPECT Beirut cTLE-RatLFP LPIRE Canine ATLE AGS IESS KAIME PCS
Benchmarks: Beirut Canine FMCE cTLE-RatLFP LPIRE HUP RESPECT AGS ATLE IESS KAIME PCS
Metrics: sensitivity F1 AUROC balanced accuracy AUPRC Brier score
Baselines: ModernTCN MRConv MultiresNet Omni-Scale SPaRCNet EEGConformer iTransformer PatchTST Pathformer SeizureFormer TSLANet Medformer
Builds on: Mamba fractional Fourier transform Weyl-Heisenberg frames linear attention TS2Vec
Closest papers by meaning
- NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.84)
- Long-Term EEG Partitioning for Seizure Onset Detection (AAAI 2025 · similarity 0.84)
- DMNet: Self-comparison Driven Model for Subject-independent Seizure Detection (NeurIPS 2024 · similarity 0.84)
- EvoBrain: Dynamic Multi-Channel EEG Graph Modeling for Time-Evolving Brain Networks (NeurIPS 2025 · similarity 0.83)
- Seizure-Semiology-Suite($S^3$): A Clinically Multimodal Dataset, Benchmark, and Models for Seizure Semiology Understanding (ICML 2026 · similarity 0.83)
- REST: Efficient and Accelerated EEG Seizure Analysis through Residual State Updates (ICML 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.