Table 3 Summary of channel selection methods
From: A review of channel selection algorithms for EEG signal processing
Technique | Subset channel | Evaluation method | Performance metrics | Application |
|---|---|---|---|---|
Selection strategy | ||||
Duun-Henriksen et al. [12] | Statistical criteria | Filtering | Sensitivity, 96 %; false detection rate, 0.14/h (with three channels) | Seizure detection |
Faul [13] | Statistical criteria | Average accuracy, 96 % on neonatal database; average accuracy, 94 % on adult database; computational effort saving, 65 % | ||
Faul and Marnane [15] | Statistical criteria | Average accuracy rate, 95.74 % (location spread), 91 % (single idle), and 91.48 % (twin idle), (with 2 channels) | ||
Atoufi et al. [16] | Sequential search | Average accuracy rate, 60 % (for EEG dataset with three channels) | ||
Shih et al. [17] | Sequential search | Wrapper | Average accuracy rate, 97 %; average detection latency, 11.2 s (with 4.6 average number of channels) | |
Glassman and Guttag [19] | Sequential search | Average false negative, 0.011; average false positive, 0.48; and average latency time, 9.54 s (with 7.1 average number of channels) | ||
Chang et al. [22] | Pre-specified | Average accuracy rate,70 % (for EEG dataset with three channels); average energy saving, 93.73 % | ||
Greene et al. [25] | Pre-specified | Average accuracy rate, 90.77 % (with single channel, C3-C4) | ||
Temko et al. [26] | Channel weighting | Average precision-recall, 84.42 (with 8 channels) | ||
Zimbric et al. [29] | Pre-specified | Human-based | Average sensitivity, 86.5 %; average specificity, 98 % (with 3 channels) | |
Tekgul et al. [30] | Pre-specified | Sensitivity, 96.8 %; specificity, 100 % (with 9 channels) | ||
He et al. [31] | Sequential search | Filtering | Average accuracy rate, ~95 % (with ~33 average number of channels) | Motor imagery Classification |
Tam et al. [34] | Sequential search | Highest average accuracy rate, 91.7 % (with 22 channels) | ||
Yong at al. [37] | Pre-specified | Average accuracy rate, 73.5 % (with 13 average number of channels) | ||
Meng et al. [39] | Heuristic algorithm | Average accuracy rate, 89.68 % (with 20 channels) | ||
Wang et al. [41] | Maximum of spatial pattern vectors | Average accuracy rate, 92.66 % (with 4 channels) and 94.96 % (with 8 channels) | ||
Shan et al. [45] | Sequential search | Accuracy rate, 63.7 (for first dataset with 2 channels) and 81.3 % (for second dataset with 16 channels) | ||
Arvaneh et al. [46] | Pre-specified | Average accuracy rate, 70.47 % (with eight channels) | ||
He et al. [47] | Genetic algorithm | Average accuracy rate, 88.2 % (for first dataset) and 89.38% (for second dataset) | ||
Arvaneh et al. [48] | Pre-specified | Average accuracy rates (SCSP1), 81.63 % (for Dataset IIa with 13.22 average channels) and 82.28 % (Dataset IVa with 22.6 average number of channels) | ||
Average accuracy rates (SCSP2), 79.07 % (for Dataset IIa with 8.55 average channels) and 79.28 % (for Dataset IVa with 7.6 average channels) | ||||
Yang et al. [51] | Pre-specified | Wrapper | Average accuracy rate, 78 % (with 11 channels) | |
Wei and Wang [53] | Random search | Accuracy rate, 83 % (for S1 with 8 channels), 91 % (for S2 with 9 channels), 75 % (for S3 with 14 channels), 86 % (for S4 with 8 channels), and 87 % (for S5 with 7 channels) | ||
Zhou and Yedida [54] | Sequential search | Average accuracy rate for healthy subjects >90 % (with 90 channels), accuracy rate for stroke subject <85 % (with 110 channels) | ||
Kamrunnahar et al. [55] | Complete search | Average classification errors, ~21.75 and 28.28 % (for subject 1 with 4 and 3 channels for tasks 1 and 2) | ||
Yang et al. [62] | Genetic algorithm | Average accuracy rate, 80 % (for 10 channels with the first dataset) and 86 % (for 6 channels with the second dataset) | ||
Lal et al. [40] | Sequential search | Embedded | Average error rate, 23 % (for 17 channels) and 24 % (12 channels) | |
Schroder et al. [65] | Sequential search | Average error rate, 26.9 % (with ≥32 channels) | ||
Li et al. [67] | Complete search | Hybrid | Average accuracy rate, 92.34 % (7 channels with dataset “aa”) and 94.63 % (8 channels with dataset “a1”) | |
Rizon et al. [68] | Pre-specified | Filtering | Minimum values of FPI = 0.150051, PME = 0.154724, and SD = 0.328312 (with 4 channels) | Emotion classification |
Jatupaiboon et al. [69] | Pre-specified | Wrapper | Accuracy rate, 84.18 % (with 5 pairs) and 85.41 % (with 7 pairs) | |
Lan et al. [72] | Sequential search | Filtering | Average accuracy rate, ~80 % (with 7, 10 channels) | Mental task classification |
Chai et al. [73] | Pre-specified | Wrapper | Accuracy rate, 65–79 % (2 channels with PSD) and 70–84 % (2 channels with HHT) | |
Tavakolian et al. [74] | Genetic algorithms | Average accuracy, 96.85 (best 6 channels’ combinations) | ||
Piryatinska et al. [75] | Complete search | Filtering | Mean agreement percentage, 87.41 % (with 4 channels) and 87.2 % (with 5 channels) | Sleep state classification |
Ong et al. [76] | Principal component analysis | Filtering | Average accuracy, 94.06 % (16 channels), 86.01 % (8 channels), and 75.13 % (4 channels) | Drug effect classification |