Automatic Seizure Detection Using Multi-Resolution Dynamic Mode Decomposition

Epilepsy is one of the most prevalent neurological issues faced by a large population around the globe. Epilepsy is marked by intermittent seizures, the detection of which can be a challenging problem. Therefore, reliably detecting the onset of seizures has evoked the interest of researchers over the last few years. Major leaps in the domain of machine learning, signal processing methods, and computational capabilities have made it a tractable task. In this paper, we apply multi-resolution dynamic mode decomposition (MRDMD), which is a data-driven dimensionality reduction technique, on the problem of epileptic seizure detection. This method can effectively separate a complex non-linear system into a collection of time-scale components at different resolutions. We have applied this algorithm on two different scalp EEG datasets, i.e., CHB-MIT and KU Leaven datasets. We have applied necessary post-processing steps to reduce the false alarm rate and boost the sensitivity and specificity. A detailed analysis of the results has been presented for the proposed method applied to both the datasets. The algorithm achieves a sensitivity of 0.937 and 0.96, a specificity of 0.99 and 0.99, a false alarm rate of 0.587 and 0.413 per hour, and a latency of 3.12 and 2.75 s for CHB-MIT and KU Leuven, respectively. The results indicate that the multi-resolution analysis yields a significant improvement in sensitivity compared with the basic dynamic mode decomposition.