Table 9.

Comparison of the accuracy (%) of the proposed model with previous works, for arousal and valence binary classification (low/high arousal, low/high valence). Values are from the original papers and using the DEAP dataset.

Year	Method	Arousal	Valence
2020	Deep Physiological Affect Network (Convolutional LSTM with a temporal loss function) [36]	79.03	78.72
2020	Attention-based LSTM with Domain Discriminator [37]	72.97	69.06
2019	Spectrum centroid and Lempel–Ziv complexity from EMD; KNN [58]	86.46	84.90
2019	Ensemble of CNNs with LSTM model [39]	—–	84.92
2019	Phase-locking value-based graph CNN [59]	77.03	73.31
2018	Time, frequency and connectivity features combined with mRMR and PCA for features reduction; Random Forest [60]	74.30	77.20
2017	Transfer recursive feature elimination; least square SVM [61]	78.67	78.75
2012	EEG Power spectral features + Asymmetry, from four bands; naive Bayes classifier [2] (DEAP paper)	62.00	57.60
2021	Proposed model	89.84	89.83
	$\alpha ,\beta ,\gamma (All-SE)+{\alpha }_{DA}(All-H3)$; KNN, K = 1