Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EEG-Titans: Long-Horizon Seizure Forecasting via Dual-Branch Attention and Neural Memory

Tien-Dat Pham, Xuan-The Tran

TL;DR

The paper tackles the challenge of forecasting epileptic seizures from EEG over long horizons, where pre-ictal signatures are subtle and traditional models struggle to retain long-range context. It introduces EEG-Titans, a memory-augmented dual-branch architecture that couples a ShallowConvNet-based spatial tokenizer with a Titans Memory-as-a-Gate backbone to efficiently model both short-term anomalies and long-range trends. Through a safety-first chronological hold-out on the CHB-MIT dataset, it achieves a segment-level sensitivity of 99.46% with an average FPR/h of 0.371, and demonstrates subject-specific adaptation (e.g., extending context to 300 s for CHB15 reducing FPR/h to 0.00 while maintaining high sensitivity). The work discusses practical deployment considerations, failure modes in artifact-heavy and pediatric data, and outlines future directions for automatic context selection, interpretability, and multimodal integration to enable trustworthy long-horizon EEG forecasting.

Abstract

Accurate epileptic seizure prediction from electroencephalography (EEG) remains challenging because pre-ictal dynamics may span long time horizons while clinically relevant signatures can be subtle and transient. Many deep learning models face a persistent trade-off between capturing local spatiotemporal patterns and maintaining informative long-range context when operating on ultralong sequences. We propose EEG-Titans, a dualbranch architecture that incorporates a modern neural memory mechanism for long-context modeling. The model combines sliding-window attention to capture short-term anomalies with a recurrent memory pathway that summarizes slower, progressive trends over time. On the CHB-MIT scalp EEG dataset, evaluated under a chronological holdout protocol, EEG-Titans achieves 99.46% average segment-level sensitivity across 18 subjects. We further analyze safety-first operating points on artifact-prone recordings and show that a hierarchical context strategy extending the receptive field for high-noise subjects can markedly reduce false alarms (down to 0.00 FPR/h in an extreme outlier) without sacrificing sensitivity. These results indicate that memory-augmented long-context modeling can provide robust seizure forecasting under clinically constrained evaluation

EEG-Titans: Long-Horizon Seizure Forecasting via Dual-Branch Attention and Neural Memory

TL;DR

The paper tackles the challenge of forecasting epileptic seizures from EEG over long horizons, where pre-ictal signatures are subtle and traditional models struggle to retain long-range context. It introduces EEG-Titans, a memory-augmented dual-branch architecture that couples a ShallowConvNet-based spatial tokenizer with a Titans Memory-as-a-Gate backbone to efficiently model both short-term anomalies and long-range trends. Through a safety-first chronological hold-out on the CHB-MIT dataset, it achieves a segment-level sensitivity of 99.46% with an average FPR/h of 0.371, and demonstrates subject-specific adaptation (e.g., extending context to 300 s for CHB15 reducing FPR/h to 0.00 while maintaining high sensitivity). The work discusses practical deployment considerations, failure modes in artifact-heavy and pediatric data, and outlines future directions for automatic context selection, interpretability, and multimodal integration to enable trustworthy long-horizon EEG forecasting.

Abstract

Accurate epileptic seizure prediction from electroencephalography (EEG) remains challenging because pre-ictal dynamics may span long time horizons while clinically relevant signatures can be subtle and transient. Many deep learning models face a persistent trade-off between capturing local spatiotemporal patterns and maintaining informative long-range context when operating on ultralong sequences. We propose EEG-Titans, a dualbranch architecture that incorporates a modern neural memory mechanism for long-context modeling. The model combines sliding-window attention to capture short-term anomalies with a recurrent memory pathway that summarizes slower, progressive trends over time. On the CHB-MIT scalp EEG dataset, evaluated under a chronological holdout protocol, EEG-Titans achieves 99.46% average segment-level sensitivity across 18 subjects. We further analyze safety-first operating points on artifact-prone recordings and show that a hierarchical context strategy extending the receptive field for high-noise subjects can markedly reduce false alarms (down to 0.00 FPR/h in an extreme outlier) without sacrificing sensitivity. These results indicate that memory-augmented long-context modeling can provide robust seizure forecasting under clinically constrained evaluation
Paper Structure (39 sections, 2 equations, 3 figures, 5 tables)

This paper contains 39 sections, 2 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Dataset and Preprocessing
  4. Dataset Description and Subject Selection
  5. Seizure Clustering and Labeling Strategy
  6. Signal Processing and Data Augmentation
  7. EEG-Titans Architecture
  8. Spatial Tokenizer - ShallowConvNet
  9. MAG Temporal Backbone
  10. Rationale
  11. Neural Memory Module
  12. MAG Block
  13. Classification Head
  14. Post-processing and Evaluation Strategy
  15. Soft Fusion Top-K and Decision Rule
  16. ...and 24 more sections

Figures (3)

  • Figure 1: Schematic of the seizure prediction protocol. The pre-ictal interval (Label 1) is defined as a 30-minute window prior to seizure onset and is separated from onset by a 5-minute Seizure Prediction Horizon (SPH). A 1.5-hour safety margin is removed between inter-ictal (Label 0) and seizure-related periods to reduce label leakage and contamination.
  • Figure 2: Overall architecture of EEG-Titans. Raw multi-channel EEG is processed by (1) a Spatial Tokenizer based on ShallowConvNet to extract spectro-spatial features and produce a token sequence; (2) a Temporal Backbone based on Titans (Memory-as-a-Gate) to model long-term dependencies via complementary local (attention) and global (memory) pathways; and (3) a classification head that outputs the seizure probability $P(t)$.
  • Figure 3: Titans: Memory-as-a-Gate (MAG) module. The module combines a local branch (causal sliding-window attention) that captures short-range patterns and a global branch (neural memory) that maintains long-range context. A learnable gate $g_t$ fuses the two representations into a context-aware output.