Epileptic seizure detection in long-term EEG records using sparse rational decomposition and local Gabor binary patterns feature extraction

In this paper, we address the problem of off-line supervised detection of epileptic seizures in long-term Electroencephalography (EEG) records. A novel feature extraction method is proposed based on the sparse rational decomposition and the Local Gabor Binary Patterns (LGBP). Namely, we decompose the channels of the EEG record into 8 sparse rational components using a group of optimal coefficients. Then, a modified 1D LGBP operator is applied, which is followed by downsampling of the data. The width of the largest LGBPs is finally computed for all the 8 rational components and the 23 channels of the EEG record. Hence, we characterize seizure patterns of one-second-long EEG epochs by 23 × 8 features. The effectiveness of the proposed feature extraction method is assessed using different classifiers which are trained with 25% of early EEG records of each patient. We performed an extensive comparative study over 163 h of EEG recordings from the CHB-MIT Scalp EEG database. The experiments show that the proposed technique outperforms other dedicated techniques by achieving the overall sensitivity of 70.4% and the overall specificity of 99.1%, in the patient-specific detection of epileptic EEG epochs. Moreover, it detects onset of seizures with the overall sensitivity of 91.13% and false alarms per hour rate of 0.35, on average.