Table 8 Feature extraction, selection methods, fusion with other biosignals for machine learning and the important features from each study
From: Machine learning methods for the study of cybersickness: a systematic review
Author | Biosignal | Feature extraction/selection methods | Feature fusion | Important features |
|---|---|---|---|---|
Biometric and neurophysiological features relevant to cybersickness | ||||
Nam et al. [12] | EEG, EOG, ECG, finger tip skin temperature, PPG, skin conductance | PCA | Yes | Fz, Cz, Pz, O1, O2, theta (5–8 Hz), alpha (9–13 Hz), beta (14–30 Hz), gamma (31–50 Hz), standard deviation of EOG, mean R-R of ECG, mean and standard deviation of fingertip skin temperature, PPG and skin conductivity |
Yu et al. [13] | EEG | PCA, LDA, NWFE, FFS/BFS for PSD | None | Delta (0.1–3 Hz), theta (4–7 Hz), alpha (8–13 Hz), beta (14–30 Hz) |
Wei et al. [16] | EEG | PCA for PSD | None | Broadband frequency 1–50 Hz |
Wei et al. [14] | EEG | Genetic algorithm for PSD | None | Broad band frequencies, especially delta (1–3 Hz), alpha (8–12 Hz), beta (13–30 Hz), channels unknown |
Ko et al. [15] | EEG | PCA for PSD | None | Fp1, Fp2, C3, C4, Pz, Oz |
Lin et al. [17] | EEG | Inheritable bi-objective combinatorial genetic algorithm (IBCGA) for PSD | None | Gamma band (21–50 Hz) (parietal area and occipital midline) |
Ko et al. [18] | EEG | Extended inheritable bi-objective combinatorial genetic algorithm (e-IBCGA) for PSD | None | Beta (13–20 Hz) and gamma (21–30 Hz) (parietal area and occipital midline) |
Lin et al. [19] | EEG | PCA for PSD | None | Alpha (8–12 Hz) and gamma (21–30 Hz) combined, broad band signals (occipital midline) |
Dennison et al. [29] | ECG, EGG, EOG, blink rate, PPG, breathing rate, GSR | Pearson correlation with SSQ cut-off | None | Bradygastric (less than 2 cycles of contraction per minute) percentage power, mean blinks, mean breaths, MSSQA |
Pane et al. [26] | EEG | ANOVA to rank frequency band feature importance based on 3 class labels (none, low, high cybersickness) | None | Decrease of Percentage power of beta (12–30 Hz) in O1 |
Mawalid et al. [21] | EEG | Mean, variation, standard deviation, number of peak and ratio logarithmic of power spectral density (power percentage) | Yes | Alpha (8–13 Hz) and beta (13–20 Hz) combined for all 14 channels, as well as their variation and standard deviation |
Khoirunnisaa et al. [20] | EEG | Channel selection through information gain and correlation-based on feature selection | None | Power percentage beta (16–32 Hz) for F3 > 01 > 02 > F4 > AF4 |
Dennison et al. [25] | EEG, ECG, EOG, blink rate, breathing rate, EGG, postural sway, head movement | Greedy sequential forward feature selection process | Yes | Number of breaths per 30 s, number of blinks per 30 s, heart rate, ECG R-peak amplitude, avatar right-left displacement, % of slow wave stomach activity (less than 2 cycles of contraction per minute), 13 EEG powerband features (0.1–30 Hz) (left frontal alpha, left motor theta, left parietal beta, left occipital delta, left occipital theta, left occipital alpha, right frontal theta, right frontal gamma, right motor delta, right motor theta, right parietal beta, right parietal delta, and right occipital gamma) |
Wang et al. [34] | Postural sway | LSTM encoder to learn features | No | Reconstruction error of postural sway signal |
Garcia-Agundez et al. [28] | ECG, EOG, blink rate, breathing rate, GSR | HR, breathing rate, respiration rate using peak detection algorithm | Yes | Combination of game content vectors, heart rate, blink rate, respiratory rate, galvanic skin response |
Jeong et al. [22] | EEG | Raw data + power bands | Yes | Signal quality weightings |
Li et al. [35] | EEG, postural sway, head body movement | PCA for Power band, centre of pressure, head and waist movement | Yes | Combination of theta (4–8 Hz) and alpha (8–13 Hz) in all 31 channels, center of pressure, head and waist movement |
Kim et al. [42] | EEG | Temporal and spectral networks | Yes | P3, P4 |
Liao et al. [27] | EEG | PSD | Yes | Broadband frequencies, 0–100 + Hz |
Li et al. [23] | EEG | Combined 4 rhythm energy ratios for all channels | None | FP1, FP2, C3, C4, P3, P4, O1, O2 |
Lee and Alamaniotis [43] | EEG | EEGNET to capture features | None | Unknown |
Islam et al. [30] | ECG, breathing rate, GSR | Pearson-correlation coefficient analysis, min, max, running average for HR, HRV and GSR | Yes | Min, max, running average for heart rate, heart rate variability and galvanic skin response |
Islam et al. [31] | ECG, breathing rate, GSR | Pearson-correlation coefficient analysis, min, max, running average for HR, HRV and GSR | Yes | Min, max, running average for heart rate, heart rate variability and galvanic skin response |
Martin et al. [33] | BVP, EDA | HRV time domain and frequency domain computation, EDA tonic and phasic feature computation | Yes | Binary and multiclassification Rank 1/50: Baseline EDA minimum amplitude Binary classification only Rank 5/50: Heart rate Multiclassification only Rank 5/50: pNN50 |
Recenti et al. [24] | EEG, EMG, heart rate | Power spectra based on previous studies | Yes | Beta EEG signals (13–35 Hz), EMG at right gastrocnemius 40–132 Hz, average HR |
Oh and Kim [32] | BVP, respiratory signal | Manual selection of HRV and respiratory signal features | Yes | HR, HRV amplitude, LF, HF, and LF/HF ratio, respiratory rate and respiratory value |