The Impact of Audio Classification on Detecting Seizures and Psychogenic Non-Epileptic Seizures

Abstract

Epilepsy vocalization feature, defined as the sound patients produce when undergoing a seizure/Psychogenic Non-Epileptic Seizure (PNES), is one of the features used to diagnose epilepsy/PNES. This study tries to analyze whether computer-aided techniques utilizing the principles of signal processing and pattern recognition can be used to classify the vocalization into epilepsy seizure or PNES. Sixteen seizure and twelve PNES samples were collected to perform the analysis. Three sound features were extracted from each sample, the maximum of the envelope and its mean, power spectral density, and Mel-Frequency Cpestral Coefficients (MFCCs). Equal test-train classification was used to determine the separability of the samples. Cross validation was then performed to confirm equal test-train findings and to analyze the efficiency of the classification using three classifiers, LDA, QDA, and SVM. Equal test-train results show that the samples are separable. Overall accuracy was 100% and true positive was 99% achieved by SVM classifier and MFCCs 4-feature space. Cross validation achieved 76% overall accuracy and 94% true positive by SVM classifier and MFCCs 4-feature space. In conclusion, it is possible to separate samples using vocalization only, however, further aspects need to be tested before generalizing the results.