[Paper Review] LookAroundNet: Extending Temporal Context with Transformers for Clinically Viable EEG Seizure Detection

Þór Sverrisson, Steinn Guðmundsson|arXiv (Cornell University)|Jan 9, 2026

EEG and Brain-Computer Interfaces0 citations

TL;DR

LookAroundNet uses a transformer to incorporate extended temporal context in EEG, achieving strong cross-dataset seizure detection with efficient inference and an ensemble that improves performance.

ABSTRACT

Automated seizure detection from electroencephalography (EEG) remains difficult due to the large variability of seizure dynamics across patients, recording conditions, and clinical settings. We introduce LookAroundNet, a transformer-based seizure detector that uses a wider temporal window of EEG data to model seizure activity. The seizure detector incorporates EEG signals before and after the segment of interest, reflecting how clinicians use surrounding context when interpreting EEG recordings. We evaluate the proposed method on multiple EEG datasets spanning diverse clinical environments, patient populations, and recording modalities, including routine clinical EEG and long-term ambulatory recordings, in order to study performance across varying data distributions. The evaluation includes publicly available datasets as well as a large proprietary collection of home EEG recordings, providing complementary views of controlled clinical data and unconstrained home-monitoring conditions. Our results show that LookAroundNet achieves strong performance across datasets, generalizes well to previously unseen recording conditions, and operates with computational costs compatible with real-world clinical deployment. The results indicate that extended temporal context, increased training data diversity, and model ensembling are key factors for improving performance. This work contributes to moving automatic seizure detection models toward clinically viable solutions.

Motivation & Objective

• Motivate automated, clinically viable seizure detection across diverse EEG datasets and settings.
• Investigate whether extending temporal context around target EEG segments improves detection accuracy.
• Evaluate generalization to unseen recording conditions and data distributions using multiple public and private datasets.
• Assess computational efficiency to ensure practicality for real-world clinical deployment.

Proposed method

• Propose LookAroundNet, which processes three-part EEG inputs (look-behind, target, look-ahead) with channel-wise transformer encoders and cross-channel attention.
• Pre-process EEG into longitudinal bipolar montage, apply filtering/resampling, and use patches per channel for transformer processing.
• Train with CrossEntropyLoss and AdamW (Schedule-Free) on 200 epochs, using label smoothing and balanced sampling across seizure, non-seizure, and mixed segments.
• Evaluate using SzCORE framework with event-based and sample-based scoring, reporting F1, sensitivity, precision, and FP/day across multiple datasets.
• Compare against prior models (EventNet, EEG-U-Transformer) and perform ablations on context window size and ensembling to assess performance gains.
• Demonstrate inference efficiency across hardware to argue clinical viability.

Experimental results

Research questions

• RQ1 Does incorporating extended temporal context around a target EEG segment improve seizure detection performance across diverse datasets?
• RQ2 How does training on data from multiple sources (public and proprietary) affect generalization to unseen recordings?
• RQ3 What is the impact of context window size and placement on detection metrics, and is ensembling multiple context configurations beneficial?
• RQ4 Can LookAroundNet achieve clinically viable inference speeds on common hardware while maintaining high accuracy?
• RQ5 What are the trade-offs between sensitivity and false positives when optimizing thresholds for clinical deployment?

Key findings

• LookAroundNet with a 16-second target window and 32+32 seconds look-behind/look-ahead achieves state-of-the-art performance across several test sets, outperforming prior models on TUSZ and Siena/SeizeIT1 while using far fewer parameters.
• Ensembling LookAroundNet variants with context windows positioned before, after, or around the target further improves event-based F1-scores and reduces false positives compared to single configurations.
• Extending the temporal context generally improves F1-scores up to about 128 seconds total context, with diminishing returns beyond that across most datasets.
• Training on a combination of public datasets (notably TUSZ plus Kvikna) yields best generalization across test sets, while adding some smaller datasets can increase false positives in certain cases.
• Inference on standard hardware runs well below real-time, with sub-6-second per-hour processing on GPUs and even faster on workstations, supporting clinical deployment.
• Contextual look-around, data diversity, and model ensembling are key factors driving improved performance and clinical viability.