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

Source: arXiv - 2601.06016v1

Overview

The paper presents LookAroundNet, a transformer‑based model that improves automated EEG seizure detection by explicitly looking at the EEG signal before and after the segment under analysis—mirroring how neurologists read EEGs in practice. Tested on a mix of public clinical datasets and a large proprietary home‑EEG collection, the approach shows strong, consistent performance while keeping inference fast enough for real‑time or bedside use.

Key Contributions

• Extended temporal context: Introduces a “look‑around” window that feeds pre‑ and post‑seizure EEG data into a transformer, capturing patterns that span several seconds to minutes.
• Transformer architecture for EEG: Adapts modern self‑attention mechanisms (originally popular in NLP and vision) to the multichannel, time‑series nature of EEG.
• Cross‑domain evaluation: Benchmarks the model on diverse datasets (routine clinical EEG, long‑term ambulatory recordings, and a large home‑monitoring cohort) to assess robustness to different hardware, patient demographics, and noise conditions.
• Efficient inference: Demonstrates that the model can run on commodity GPUs/edge devices with latency suitable for bedside alerts.
• Ablation insights: Shows that performance gains stem mainly from (1) the wider temporal window, (2) training on heterogeneous data, and (3) simple model ensembling.

Methodology

1. Data preprocessing – Raw EEG is band‑pass filtered (0.5–70 Hz), re‑referenced, and split into 2‑second epochs. For each target epoch, a symmetric context window (e.g., ±10 s) is concatenated, giving the model a “look‑around” view.
2. Model architecture –
   • Patch embedding: Each channel’s time‑series within the window is linearly projected into a low‑dimensional token.
   • Positional encoding: Temporal order is encoded so the transformer can distinguish “past” vs. “future” tokens.
   • Transformer encoder: Stacked self‑attention layers learn relationships across channels and across time, allowing the network to spot subtle pre‑ictal or post‑ictal signatures.
   • Classification head: A lightweight feed‑forward network maps the pooled transformer output to a binary seizure / non‑seizure decision.
3. Training strategy – The authors train on a pooled dataset that mixes multiple public sources and the proprietary home‑EEG set, using standard cross‑entropy loss and data augmentation (channel dropout, Gaussian noise).
4. Ensembling – At inference, predictions from three independently trained instances are averaged, yielding a modest boost in robustness without a large computational penalty.

Results & Findings

• Consistent performance: LookAroundNet outperforms baseline CNN‑only models by 4–7 % absolute F1 across all datasets.
• Generalization: When trained on the public data only and tested on the home‑EEG set, the drop in F1 is <3 %, indicating good transferability.
• Latency: Average inference time ≈ 30 ms per 2‑second epoch on an NVIDIA Jetson Xavier, well within real‑time alerting requirements.
• Ablation: Removing the context window reduces F1 by ~5 %; training on a single dataset reduces it by ~6 %; ensembling adds ~1–2 % gain.

Practical Implications

• Real‑time bedside monitoring: Hospitals can deploy LookAroundNet on existing EEG acquisition hardware to generate instant seizure alerts, potentially reducing the workload of neurophysiologists.
• Home‑based tele‑monitoring: The low‑latency, edge‑friendly design makes it feasible to embed the model in portable EEG headsets, enabling continuous seizure surveillance for patients with refractory epilepsy.
• Data‑driven protocol design: By showing that diverse training data improves robustness, the work encourages manufacturers and clinics to share anonymized EEG recordings, accelerating the creation of universal seizure detectors.
• Integration with clinical workflows: The model’s output can be combined with visual review tools (e.g., highlighting suspicious windows) to create a “human‑in‑the‑loop” system that speeds up diagnosis while preserving expert oversight.

Limitations & Future Work

• Label quality: The study relies on expert annotations that may vary across sites; noisy labels could limit ceiling performance.
• Window size trade‑off: Larger look‑around windows improve detection but increase memory usage; finding the optimal balance for ultra‑low‑power devices remains open.
• Explainability: While attention maps provide some insight, the model is still a black box for clinicians; future work should explore more interpretable visualizations.
• Broader modalities: Extending the approach to multimodal data (e.g., video, ECG) could further boost detection accuracy, especially for subtle seizures.

Bottom line: LookAroundNet demonstrates that borrowing the “context‑aware” reasoning of human experts—and coupling it with modern transformer architectures—can bring automated seizure detection a step closer to everyday clinical and home use.

Authors

• Þór Sverrisson
• Steinn Guðmundsson

Paper Information

• arXiv ID: 2601.06016v1
• Categories: cs.LG
• Published: January 9, 2026
• PDF: Download PDF