A SPCNN Model for Patient-Independent Prediction of Epilepsy Using MFCC Features

Abstract

Epilepsy is one of the most common psychiatric disorders in humans, and the sudden onset of seizures can seriously affect patients' lives. Predicting seizures can help prevent accidents and help physicians to intervene in treatment. Most studies on seizure prediction have chosen to customize prediction models for patients for high accuracy and sensitivity, which are difficult to adapt to the high variability between electroencephalogram (EEG) signals of different patients and cannot be applied to other patients and are difficult to use clinically. The main energy of EEG signal is concentrated in the low-frequency phase, which contains more detailed information, inspired by some methods in speech signal processing. The SPCNN, a patient-independent epilepsy prediction model, was constructed using convolutional neural networks by introducing more Mel-Frequency Cepstral Coefficients (MFCC) features concentrated in the low-frequency region, and obtained 93% accuracy, 91 % sensitivity, and 83% F1-score values in the CHB-MIT dataset.