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

TL;DR

LookAroundNet addresses the need for robust, clinically viable EEG seizure detection by integrating extended temporal context through a transformer-based architecture. By processing a target EEG segment together with a configurable look-behind and look-ahead context, it captures longer-range temporal dependencies that mirror clinician interpretation. The method demonstrates strong cross-dataset generalization across public and proprietary ambulatory data, with ensembling of multiple context configurations achieving state-of-the-art event-based performance while maintaining practical inference costs. The work highlights the importance of diverse training data, temporal-context modeling, and pragmatic evaluation via SzCORE for moving automated seizure detection toward real-world clinical deployment. It also outlines future directions for handling artifacts, improving short-seizure detection, and enabling edge-ready implementations.

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.

Figures (7)

• Figure 1: The LookAroundNet architecture. The input window includes a target segment, which is the specific EEG segment to be classified, along with surrounding context segments that provide additional temporal information. Each input channel is independently patched, and features are extracted from each patch. A transformer encoder processes each channel, followed by multi-head attention across channels and a classification layer.
• Figure 2: Test set performance of selected models and training set combinations, shown as seizure detection rate versus false positives per day. Each subfigure corresponds to a different test set, with points representing individual model–training set combinations.
• Figure 3: Comparison of event-based F1-scores across test datasets for different context window configurations, each with a total extended context duration of 64 seconds. The highlighted area corresponds to the target segment, and the size and positions of the colored bars represent the context window size and placement. A negative context size indicates past context, whereas a positive size indicates future context. The dashed line represents the ensembled model, created by combining outputs from the three look-around configurations.
• Figure 4: Relationship between event-based F1-score and total context window duration across test sets. Each plot shows results for a model using equal context durations before and after a 16-second target segment.
• Figure 5: Age distribution of patients in the training, validation, and test sets.