.png){kind=icon}
.png){kind=icon}
Return
DiEff-MSwinNet: Dilated efficientnet based multi-scale swin transformers for spectrogram-powered epileptic seizure detection in real time
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.114927.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate epileptic seizure detection and prediction are crucial for timely clinical intervention and improved patient outcomes. Existing deep learning models face challenges in handling EEG signal complexity due to patient-specific seizure variations, noise, artifacts, insufficient fine-grained time–frequency feature extraction, high false alarm rates, and limited generalization across diverse patients. To address these limitations, we propose DiEff-MSwinNet, a novel deep learning framework for real-time seizure detection and early prediction. It incorporates Adaptive Artifact Suppression (AAS) to dynamically remove noise while preserving seizure-relevant features, and the Fractional Stockwell Transform (FrST) enhances spectral-temporal resolution, converting EEG signals into detailed spectrograms that reveal subtle seizure indicators. For feature extraction, Dilated EfficientNet (DiEff) captures low-amplitude seizure patterns via dilated convolutions, while Multi-Scale Swin Transformer Network (MSwinNet) models long-range temporal dependencies efficiently using multi-scale attention. A hybrid classification head leveraging temporal modelling and self-attention refines feature representations, reducing false alarms and improving prediction precision. A probabilistic decision mechanism differentiates pre-ictal, inter-ictal, and ictal states. Additionally, Dynamic Gradient Alignment with Contrastive Meta-Learning (DGA-CM) ensures cross-patient generalizability by learning patient-invariant features. Validated on CHB-MIT and TUH EEG datasets, DiEff-MSwinNet achieves accuracies of 99.11 % and 99.62 %, sensitivities of 98.92 % and 99.03 %, and AUCs of 0.9845 and 0.9908, respectively. It predicts seizures 5–15 min before onset (12.3 min for CHB-MIT, 11.8 min for TUH) with 96.8–97.5 % reliability, enabling proactive intervention. These results establish DiEff-MSwinNet as a clinically reliable, high-performance solution for real-time seizure detection and early prediction.
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W
View more organizations