Investigating music algorithm in neural information processing: robust feature extraction for emotional state classification from multichannel EEG signals

Abstract

The major challenge in any electroencephalogram (EEG) classi cation task lies with in the dilemma of feature extraction, as raw time series signal provide little correlated information, yet it holds colossal varieties of hidden feature patterns. Frequency domain transformations are considered state of the art to tackle such complexity. Nevertheless such conventional feature extraction techniques for instance; discrete wavelet transformation (DWT), short time Fourier transform (STFT), di erential entropy or classical non-parametric power spectral density (PSD) estimation models are computationally expensive as they demonstrate high computational complexity and extensive run time. Consequently arti cial intelligence driven multi channel EEG based systems struggles to process neural information in real time and such barrier minimizes the dynamics of relevant human computer interaction (HCI) and brain computer interaction (BCI) applications. Multiple Signal Classi cation Algorithm (MUSIC) is an eigen decomposition based parametric PSD estimation model, which solely uses linear transformation rather than computing windowed periodigram from autocorrelted function of the targeted signal for transformation. Hence MUSIC algorithm should demonstrate lesser time complexity and run time than contemporary classical non-parametric PSD models. Nevertheless this particular model is relatively unexplored for such feature extraction task speci cally in the area of emotion recognition, as the model is di cult to implement in terms of EEG signals which demonstrate random behaviour. Our research investigates the performance of MUSIC algorithm in feature extraction task for emotion recognition from multi channel EEG signals and compares its performance with conventional classical non parametric models. It also clari es the complexity in subspace estimations for EEG waveforms through in detailed analysis, which are indispensable parameters for implementing any eigen decomposition based models in such particular cases. Our proposed model derived state of the art 5-fold cross validation accuracies of 97% and 97:4% for Multi Layer Perceptron (MLP) network and Hybrid Long Term Short Memory (LSTM)-MLP network, respectively on the SEED emotional dataset. The proposed MUSIC model optimizes 95%􀀀96% run time comparing with conventional classical non-parametric techniques for feature extraction. With exceptional 0:01 sec: machine speci c run time for feature extraction task, the proposed model shows great prospect in real time applications. Network performance and advanced visualization techniques demonstrate the MUSIC model based feature space holds signi cant superiority over non-parametric model generated feature space. Additionally the research also found extensive aws in the widely popular SEED dataset, which were ignored in previously. Over 17% trials were found to hold multiple corrupt channel resulted from external artifacts, which should have e ected previously conducted researches. Our research also discusses the e ects of such awed trial in network performance.