Effects of Data Augmentation with the BNNSMOTE Algorithm in Seizure Detection Using 1D-MobileNet

Table 8

Comparison of multiple approaches’ experimental results on the CHB-MIT dataset.

Authors	Year	Methods	Patients	Rec (%)	Pre (%)	F1 (%)	Acc (%)
Khan et al. [34]	2012	Multiple wavelet scales, LDA	5	83.6	86.7	85.1	91.8
Janjarasjitt [35]	2017	Wavelet based features, SVM	24	72.99	—	—	96.87
Yao et al. [36]	2018	Attention, BiLSTM	23	87.30	88.29	87.74	87.80
Wei et al. [37]	2019	MIDS, WGANs, 1D-CNN	24	72.11	—	—	84.00
Yao et al. [38]	2019	Windowing, IndRNN	24	88.80	88.69	88.71	88.70
Yuan et al. [39]	2019	STFT-mConvA	23	85.00	85.68	85.34	94.34
Hu et al. [40]	2020	K-means SMOTE, RFS + extra-trees + GBDTs	22	—	—	85.81	89.49
Jiang and Zhao [21]	2021	SMOTE + TomekLink, CSL + SVM	22	86.34	—	—	94.00
Boonyakitanont et al. [41]	2021	CNN, ScoreNet	24	76.54	64.74	70.15	99.83
Ryu and Joe [42]	2021	DWT, DenseNet-LSTM	24	92.92	91.71	92.30	93.28
Zhao et al. [43]	2021	GAT, focal loss	23	97.10	99.59	98.33	98.89
Guo et al. [44]	2022	Isolation forest, time-frequency feature, and easyensemble	24	95.55	—	—	92.62
This work	2022	BNNSMOTE, 1D-MobileNet	20	87.46	97.17	91.90	99.40

LDA: linear discriminant analysis, MIDS: merger of the increasing and decreasing sequences, WGANs: Wasserstein generative adversarial nets, IndRNN: independently RNN, mConvA: multiconvolutional autoencoder, GBDTs: gradient boosting decision trees, CSL: cost-sensitive learning, DWT: discrete wavelet transform, and GA T: graph attention network.