Residual and bidirectional LSTM for epileptic seizure detection

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epileptic seizures affect over 70 million people globally.
• Visual interpretation of electroencephalogram (EEG) signals for epilepsy detection is challenging and time-consuming.
• Automated methods are needed to improve the efficiency and accuracy of epilepsy diagnosis.

Purpose of the Study:

• To introduce a novel hybrid deep learning model, ResBiLSTM, for automated epileptic seizure detection using EEG signals.
• To evaluate the performance of ResBiLSTM on established epilepsy datasets.
• To compare ResBiLSTM against state-of-the-art deep learning approaches.

Main Methods:

• A hybrid deep learning architecture combining a one-dimensional residual neural network (ResNet) for feature extraction and a bidirectional long short-term memory (BiLSTM) layer for temporal dependency modeling.
• ResNet extracts local spatial features from EEG signals.
• BiLSTM processes these features to capture temporal dynamics, followed by fully connected layers for classification.

Main Results:

• ResBiLSTM achieved high accuracy rates of 98.88-100% in binary and ternary classifications on the University of Bonn dataset.
• On the Temple University Hospital (TUH) seizure corpus, ResBiLSTM attained 95.03% classification accuracy and a 95.03% weighted F1 score for seven epilepsy seizure types.
• The model demonstrated superior performance compared to several recent state-of-the-art deep learning methods.

Conclusions:

• The ResBiLSTM model offers a straightforward and efficient solution for automated epileptic seizure detection.
• The hybrid deep learning approach effectively leverages both spatial and temporal features in EEG signals.
• ResBiLSTM shows significant promise for improving the diagnosis and management of epilepsy.