An Adaptive Structure Neural Networks with Application to EEG Automatic Seizure Detection

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Multi-layer back-propagation networks often use fixed structures, limiting adaptability and performance.
• Neuron pruning strategies can lead to information loss, while neuron generation strategies are less explored.

Purpose of the Study:

• To introduce a novel neuron-generating algorithm for adaptive multi-layer perceptron networks.
• To enhance network performance, generalization capabilities, and efficiency in training.
• To apply and validate the algorithm in the critical domain of automatic epileptic seizure detection.

Main Methods:

• A neuron-generating strategy is employed, contrasting with neuron-pruning approaches.
• Network structure adaptively adjusts by generating new neurons based on stabilized error.
• New neurons are strategically placed to address areas contributing most to network error.

Main Results:

• The proposed adaptive network architecture demonstrates improved performance and generalization over fixed-structure networks.
• Application to electroencephalogram (EEG) data for epileptic seizure detection shows accurate identification of all true seizures.
• The algorithm achieves a 60-70% reduction in training epochs and reduces false seizure detections to zero.

Conclusions:

• The novel neuron-generating algorithm offers significant advantages in network training efficiency and accuracy.
• The adaptive structure enhances generalization capabilities, crucial for complex pattern recognition tasks.
• This approach shows high efficacy in real-world applications like automated medical diagnosis, specifically epileptic seizure detection.