Power complexity feature-based seizure prediction using DNN and firefly-BPNN optimization algorithm

Epileptic seizure prediction is an online clinical application for pediatric patient monitoring. In this paper, we have introduced a novel method for detecting and predicting the seizure attack. After signal preprocessing, the time and frequency domain features are extracted. In our scenario, by estimating the power spectrum using time samples of windowed signal, the features such as non-linearity model and complexity of power for demonstrating the signal behavior are extracted. Our complexitybased feature is named Power Complexity Feature (PCF). The optimal features are selected by a hybrid model of Firefly optimization algorithm (FA) and Back Propagation Neural Network (BPNN). With these features, initial optimized MLP is trained in offline mode. A Dynamic Neural Network (DNN) based on Non-Auto Regressive (NAR) architecture estimates the EEG signal. With the trained classifier in offline mode, the predicted signals with optimal features are recognized in two classes. The initial classifier in each training stage is updated. Also, the initial dead part of signal and length of prediction by Monte-Carlo analysis and considering a similarity criterion are improved. Ultimately, the seizure signals by optimized features are recognized with accuracy rate of 86.8% in offline mode and also accuracy rate of 85.7% for the predicted signal with prediction time of 3.12 seconds is obtained.