FAPEX: Fractional Amplitude-Phase Expressor for Robust Cross-Subject Seizure Prediction

Ruizhe Zheng§Research Institute of Intelligent Complex Systems, Fudan University [email protected], Lingyan Mao§Department of Neurology, Zhongshan Hospital, Fudan University [email protected], Tian LuoChildren’s Hospital of Fudan University [email protected], Yi WangChildren’s Hospit · 2025

Read FAPEX: Fractional Amplitude-Phase Expressor for Robust Cross-Subject Seizure Prediction by Ruizhe Zheng§Research Institute of Intelligent Complex Systems, Fudan University [email protected], Lingyan Mao§Department of Neurology, Zhongshan Hospital, Fudan University [email protected], Tian LuoChildren’s Hospital of Fudan University [email protected], Yi WangChildren’s Hospit (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-03263.txt.

Introduction

Ruizhe Zheng and co-authors build FAPEX, a system meant to warn of epileptic seizures across people it has never seen. Brain signals differ a lot from person to person and across recording setups, which makes a single warning model hard to trust. FAPEX uses a learnable frame that splits signals by both time and pitch, with fractional math that keeps high and low parts in balance.

From that split, the system reads instant strength and phase patterns that may mark the lead-up to a seizure. Scores rose by about a tenth over strong rivals on standard tests. The design also aims to hold up on new kinds of recordings. Better cross-person warnings could give patients and carers precious minutes to prepare.

Catalog record

CREATORRuizhe Zheng§Research Institute of Intelligent Complex Systems, Fudan University [email protected], Lingyan Mao§Department of Neurology, Zhongshan Hospital, Fudan University [email protected], Tian LuoChildren’s Hospital of Fudan University [email protected], Yi WangChildren’s Hospit
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE44 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-03263

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

FAPEX: Fractional Amplitude-Phase Expressor for Robust Cross-Subject Seizure Prediction Ruizhe Zheng§Research Institute of Intelligent Complex Systems, Fudan University [email protected], Lingyan Mao§Department of Neurology, Zhongshan Hospital, Fudan University [email protected], Tian LuoChildren’s Hospital of Fudan University [email protected], Yi WangChildren’s Hospital of Fudan University [email protected], Dingding Han School of Information Science and Technology, Fudan University [email protected], Jing Ding*Zhongshan Hospital, Fudan University [email protected], Yuguo Yu*State Key Laboratory of Brain Function and Disorders and MOE Frontiers Center for Brain Science,Research Institute of Intelligent Complex Systems and Institutes of Brain Science, Fudan University,Shanghai Artificial Intelligence Laboratory, Shanghai 200232, China [email protected] Abstract Precise, generalizable subject-agnostic seizure prediction (SASP) remains a fundamental challenge due to the intrinsic complexity and significant spectral variability of electrophysiological signals across individuals and recording modalities. We propose FAPEX , a novel architecture that introduces a learnable fractional neural frame operator (FrNFO) for adaptive time-frequency decomposition. Unlike conventional models that exhibit spectral bias toward low frequencies, our FrNFO employs fractional-order convolutions to capture both high and low-frequency dynamics, achieving approximately 10 % 10\% improvement in F1-score and sensitivity over state-of-the-art baselines. The FrNFO enables the extraction of instantaneous phase and amplitude representations that are particularly informative for preictal biomarker discovery and enhance out-of-distribution generalization. FAPEX further integrates structural state-space modeling and channelwise attention, allowing it to handle heterogeneous electrode montages. Evaluated across 12 benchmarks spanning species (human, rat, dog, macaque) and modalities (Scalp-EEG, SEEG, ECoG, LFP), FAPEX consistently outperforms 23 supervised and 10 self-supervised baselines under nested cross-validation, with gains of up to 15 % 15\% in sensitivity on complex cross-domain scenarios. It further demonstrates superior performance in several external validation cohorts. To our knowledge, these establish FAP(O)EX as the first epilepsy model to show consistent superiority in SASP, offering a promising solution for discovering epileptic biomarker evidence supporting the existence of a distinct and identifiable preictal state and clinical translation. 1 Introduction Epilepsy is a common, heterogeneous set of neurological disorders characterized by recurrent, hypersynchronous discharges that disrupt normal cognition and behavior. Affecting over 50 million people worldwide, its diagnosis and monitoring rely fundamentally on electrophysiological recordings—whether invasive (e.g., electrocorticography (ECoG), stereo-electroencephalography (SEEG), local field potential (LFP)) or non-invasive (scalp EEG). Although seizures have long been viewed as abrupt and unpredictable events, a growing body of work demonstrates the existence of a preictal stage marked by subtle neural and behavioral changes, offering an actionable window for intervention. Figure 1 : Summary of our work. (a) Definition of the different brain activity stages for the predictive analysis of epileptic seizure (a scalp-EEG record is shown for the purpose of illustration). (b) Overview of the FAPEX development and validation pipeline. (c) Comparative evaluation results demonstrate that FAPEX consistently outperforms state-of-the-art (SOTA) supervised and self-supervised approaches across 12 diverse benchmarks in terms of F1 and sensitivity, demonstrating superior performance and generalization. Seizure prediction systems seek to detect these preictal alterations and raise alarms sufficiently in advance. As shown in Fig. 1 (a), within the established framework of ictogenesis - which delineates interictal, ictal, and postictal phases - the preictal interval offers a crucial target for clinical interventions ranging from simple alerts aimed at mitigating injury risk to sophisticated closed-loop neuromodulation devices. To formalize practical deployment, seizure prediction systems are typically evaluated with respect to two time parameters: the Seizure Prediction Horizon (SPH), which defines the minimum interval between a raised alarm and seizure onset to allow meaningful intervention, and the Seizure Occurrence Period (SOP), a predefined window during which a seizure is expected following an alarm. Why subject-agnostic seizure prediction (SASP)? Despite remarkable advances in seizure prediction achieved by pioneering studies, the field remains constrained by two fundamental limitations: the reliance on subject-specific modeling paradigms and limited scalability. Subject-specific approaches, while often achieving impressive performance on individual patients, require extensive labeled data collection for each new patient and cannot leverage knowledge across diverse patient populations. This impedes large-scale clinical adoption and negates the potential advantages of aggregating data to identify generalizable seizure biomarkers. Beyond subject specificity, additional obstacles include narrow EEG modality ranges, inconsistent preprocessing pipelines, and dependence on rigid electrode configurations. Together, these factors highlight the urgent need for truly subject-agnostic predictive algorithms capable of operating robustly across various patient and recording configurations. Specific challenges include: (1) Capturing refined high- and low-frequency biomarkers. Clinical evidence shows pathological high-frequency oscillations and low-frequency fluctuations serve as crucial epileptogenesis biomarkers. These subtle, non-stationary features are easily obscured by artifacts. Conventional CNNs and Transformers exhibit spectral biases toward low frequencies, struggling to preserve transient HFO signatures [ 63 ] . (2) Modeling phase–amplitude interactions. Epilepsy exhibits abnormal phase-amplitude coupling and (de)synchronization during seizure initiation and propagation across frequency bands. These interactions provide critical clues for distinguishing ictal from interictal states. Current clinical models typically utilize amplitude information in time or frequency domains separately, rarely integrating both. Our architecture captures these fundamental aspects of neural oscillations, leveraging their complementary insights into seizure evolution. Figure 2 : Interpretability of FAPEX . (a) Kernel density estimates of power spectral density (PSD) responses for FrNFO filters across layers and brain frequency subbands. As depth increases, the operator progressively refines its discrimination among subbands, maintaining the natural low-frequency, high-energy and high-frequency, low-energy distribution, with energy gradually stabilizing after intermediate layers. (b) Layer-wise frequency-specific gain relative to the initial layer. Unlike non-fractional operators, FrNFO consistently amplifies both low- and high-frequency components, achieving balanced cross-frequency representations, indicating its ability to capture both fast and slow neural dynamics essential for seizure prediction. (3) Handling heterogeneous channel layouts. Seizure onset zones and preictal activity distribution vary significantly across individuals, with predictive features appearing on different electrodes from patient to patient. Implantation strategies, montage configurations, and regional coverage introduce further variability in channel characteristics. Naively pooling signals across channels obscures patient-specific biomarkers and amplifies noise, compromising generalization in subject-agnostic contexts. Present work. To overcome these challenges, we propose FAPEX , a unified model for effective generalization across heterogeneous EEG settings, electrode configurations, and clinical subtypes. Our approach integrates three key innovations: (1) fractional neural frame operator (FrNFO): a learnable bank of Weyl-Heisenberg filters for adaptive time-frequency decomposition. FrNFO extracts high-fidelity features through fractional-order convolutions with minimal spectral leakage, capturing both high and low-frequency components of epileptic signals. (2) amplitude-phase cross-encoding (APCE): A bidirectional state-space architecture processing phase and amplitude representations, learning time-varying relationships to extract seizure evolution patterns. (3) Spatial correlation aggregation (SCA): Channel-wise attention mechanisms modeling inter-electrode dependencies to identify predictive spatial patterns. Together, these components enable FAPEX to learn multi-scale representations that capture phase-amplitude coupling while handling non-stationarity and channel heterogeneity. Together , these components enable FAPEX to learn rich, multi-scale representations that capture subtle changes in phase-amplitude coupling across frequencies while adaptively handling non-stationarity and channel heterogeneity. As illustrated in Fig. 2, Fr NFO serves as the foundational component, addressing low-frequency bias by preserving fragile yet critical high-frequency oscillations while providing fine-grained decomposition of amplitude and phase features for a more comprehensive picture of neural activity in seizure prediction. Our main innovations are: (1) FAPEX , a subject-agnostic framework integrating our novel Fractional Neural Frame Operator (Fr NFO ), amplitude-phase cross-encoding, and spatial correlation aggregation to anticipate seizures across diverse modalities. (2) Fr NFO , a learnable bank of Weyl–Heisenberg filters that performs fractional-order time-frequency decomposition,

Read the complete text