2026 American Academy of Neurology Abstract Website

Objective:

To develop a multitask learning framework for robust interictal epileptiform discharge (IED) detection, enabling the joint modeling of IED and sleep-related features within a unified network, leading improved performance.

Background:

Epilepsy is a common neurology disorders. IED plays a critical role in the diagnosis of epilepsy. However, two major challenges remain: (1) sleep-related physiological waveforms often resemble IEDs, leading to frequent misclassifications; and (2) conventional single-task models struggle to effectively incorporate sleep information to improve discriminative capability. As a result, these models struggle to disentangle IEDs from confounding sleep-induced patterns, resulting in poor generalization across patients and recording conditions.

Design/Methods:

The proposed network termed of Siamese-ES is designed under the hard parameter sharing paradigm, and consists of three main components: (1) electrical signal preprocessing, (2) a shared Frequency-Temporal Converter, and (3) two task-specific branches—Twins-Electron and Twins-Sleep. Siamese-ES was developed and validated on EpiSet-265K, a large-scale benchmark dataset comprising 265,551 labeled samples with both IED and sleep-related annotations. Several classic networks for EEG analysis, including EEGNet, CNN-LSTM, VGG-16, EEGFormerNet, AiED, nEpiNet, were applied for performance comparisons.

Results:

Extensive experiments demonstrate that Siamese-ES achieves state-of-the-art performance, reaching 98.46% specificity and 71.18% precision at 82.11% sensitivity, superior the best results of previous networks by 0.3% and 5%, respectively. The model achieves an AUROC of 0.978 and AUPRC of 0.819. Ablation study demonstrated the effectiveness of sleep feature integration and multi-task learning, increasing F1 score by 0.71% and 1.29%, respectively.

Conclusions:

The present study demonstrated that multi-task learning and sleep-aware feature integration capture complementary patterns between multi-channel electrophysiological signals, therefore significantly improve overall detection performance. This study lays a solid foundation for developing clinically deployable deep learning-based IED detection systems, with strong potential to enhance diagnostic accuracy, robustness, and reliability in real-world clinical settings.