Machine learning based novel cost-sensitive seizure detection classifier for imbalanced EEG data sets

Abstract

Epilepsy is one of the most prevalent neurological disorders. Its accurate detection is a challenge since sometimes patients do not experience any prior alert to identify a seizure. Electroencephalography (EEG) recordings are used for seizure detection, but these are usually of longer duration, and as a result, the behavior of the inherent data set is highly imbalanced. To detect seizures in such a scenario is a challenging task; using a typical classifier such as decision tree and decision forest can result in highly skewed class value (non-seizure), causing incorrect detection of epileptic patients. To solve this, a cost-sensitive learning method with a random forest was used. An algorithm that helps in seizure detection by penalizing the cost of a false negative concerning the duration of an EEG recording was proposed. The experimental results show that executing the classifier without penalty or inadequate penalties to the cost matrix is not a satisfactory solution. As a result, the algorithm provides up to 100% recall, which means all the seizure seconds are detected. The proposed method substantiates achieving higher actual seizure detection rates; the imposed penalty should be equal to the time duration of the EEG recordings (in seconds) for a patient. Hence, it can be potentially applied to the pre-consultation to the neurologist at the Outpatient Department for the actual seizure detection cases and refer them to the neurology department for further consultation.