.png){kind=icon}
.png){kind=icon}
Return
AttenEpilepsy: A 2D convolutional network model based on multi-head self-attention
.png){kind=icon}
PRE
AI
DOI：10.1016/j.enganabound.2024.105989.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The existing epilepsy detection models focus more on local information than the true meaning of long-range dependence when capturing time-frequency image features. This results in imprecise feature vector extraction and room for optimization of detection accuracy. AttenEpilepsy is a novel 2D convolutional network model that uses a multi-head self-attention mechanism to classify epileptic seizure periods, inter-seizure periods, and health states of single-channel EEG signals. The AttenEpilepsy model consists of two parts, namely feature extraction and time-frequency context encoding (STCE). A feature extraction method combining multi-path convolution and adaptive hybrid feature recalibration is proposed, in which multi-path convolution with convolution kernels of different sizes is used to extract relevant multi-scale features from time-frequency images. STCE consists of two modules: multi-head self-attention and causal convolution. A modified multi-head self-attention mechanism is used to model the extracted time-frequency features, and causal convolution is used to analyse the frequency information on the time dependencies. A public dataset from the University of Bonn Epilepsy Research Center is used to evaluate the performance of the AttenEpilepsy model. The experimental results show that the AttenEpilepsy model achieved accuracy (AC), sensitivity (SE), specificity (SP), and F1 score (F1) of 99.81%, 99.82%, 99.89%, and 99.83%, respectively. Further testing of the robustness of the model is conducted by introducing various types of noise into the input data. The proposed AttenEpilepsy network model outperforms the state-of-the-art in terms of various evaluation metrics.
Keywords:
Long-range dependence
Time-frequency image
Feature extraction
Time-frequency context encoding
Multi-head self-attention
Causal convolution
Journal
IF:
4.1
Papers:
5.8K
Citations:
9.4K