[Paper Review] EEG-Titans: Long-Horizon Seizure Forecasting via Dual-Branch Attention and Neural Memory

Tien-Dat Pham, Xuan-The Tran|arXiv (Cornell University)|Jan 20, 2026

EEG and Brain-Computer Interfaces0 citations

TL;DR

EEG-Titans introduces a dual-branch architecture with neural memory to forecast seizures over long horizons from EEG, achieving high segment-level sensitivity with a safety-focused, causal evaluation on CHB-MIT data.

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

Motivation & Objective

• Motivate accurate seizure forecasting from EEG by capturing long-term pre-ictal dynamics.
• Develop a memory-augmented architecture that combines short-term anomalies with long-range context.
• Evaluate under a safety-first chronological hold-out protocol to preserve temporal causality.
• Analyze failure modes on artifact-heavy and pediatric EEG to guide personalization and robustness.

Proposed method

• Propose a dual-branch architecture with a spatial tokenizer and a Titans-based temporal backbone with neural memory (MAG).
• Use ShallowConvNet as a spatial tokenizer to produce token sequences from multi-channel EEG segments.
• Employ a Memory-as-a-Gate (MAG) temporal backbone that fuses local causal attention with a neural memory pathway via a learnable decay gate.
• Train with a window-based soft fusion and per-subject thresholding to manage variability and reduce false alarms (30-minute refractory period).
• Evaluate using segment-based sensitivity and false-positive rate per hour under a chronological hold-out protocol to reflect clinical deployment.

Experimental results

Research questions

• RQ1 Can memory-augmented temporal modeling improve long-horizon pre-ictal forecasting on CHB-MIT EEG data?
• RQ2 Does combining local attention with a neural memory pathway reduce false alarms while maintaining high sensitivity under causal evaluation?
• RQ3 How does subject-specific adaptation (context extension) affect performance on artifact-heavy or non-stationary pediatric EEG?
• RQ4 What are the limitations and failure modes of memory-augmented models in pediatric EEG with varying artifact loads?

Key findings

• EEG-Titans achieves 99.46% average segment-level sensitivity with 0.3713 FPR/h across 18 CHB-MIT subjects under chronological hold-out.
• 16 of 18 subjects show 100% sensitivity with low FPR/h (<0.2 alarms/hour) in the baseline configuration.
• Subject CHB15 shows a high baseline FPR/h (3.8710) which drops to 0.00 FPR/h when context is extended from 60 s to 300 s, demonstrating practical adaptation.
• Subject CHB06 is highly sensitive (100%) but has elevated FPR/h (0.90) likely due to pediatric-age-specific non-stationarity and brief seizures, indicating need for pediatric calibration.
• Ablation studies show memory-augmented TITANS lowers FPR/h compared to attention-only and LSTM baselines, supporting a long-context stabilizing effect.
• Compared to state-of-the-art CHB-MIT results, EEG-Titans (Memory) attains competitive 99.46% sensitivity with modest FPR/h, using tokenization-based features without explicit TF transforms.