From: Machine learning methods for the study of cybersickness: a systematic review
Authors | Biosignal | Algorithm | Binary/multiclass | Accuracies | Classification type |
---|---|---|---|---|---|
Machine learning models | |||||
Nam et al. [12] | EEG, EOG, ECG, finger tip skin temperature, PPG, skin conductance | ANN, 2-layer feedforward neural network | Binary | Minimum mean square error 0.092 | Detection |
Yu et al. [13] | EEG | GMLC, KNN, SVM | Binary | KNN, NWFE 99.9% | Detection |
Wei et al. [16] | EEG | RBFNN, SVR, LR | Multi class motion sickness level | 84.07% LR 84.75% RBFNN, 86.92% SVR | Detection |
Wei et al. [14] | EEG | RBFNN | Multi class motion sickness level | 84.39% ± 0.75 | Detection |
Ko et al. [15] | EEG | LR, PCR | Multi class motion sickness level | PCR 78.3% ± 8.0 LR 64.7% ± 15.6 | Detection |
Lin et al. [17] | EEG | SVM | Multi class motion sickness level | 36.3–73.3% | Detection |
Ko et al. [18] | EEG | SVM | Multi class motion sickness level | 58.5–97.0% | Detection |
Lin et al. [19] | EEG | SONFIN, LR, SVR | Multi class motion sickness level | Broad band EEG SONFIN 82% ± 2 SVR 79% ± 3 LR 80% ± 3 | Detection |
Dennison et al. [29] | ECG, EGG, EOG, blink rate, PPG, breathing rate, GSR | Stepwise regression | SSQ score estimation | Adjusted R2: Cybersickness 0.296 Nausea 0.101 Oculomotor 0.674 Disorientation 0.268 | Detection |
Pane et al. [26] | EEG | CN2 rule induction algorithm, decision tree, SVM | Multiclass | CN2 88.9% Decision tree 72.2% SVM 83.3% | Detection |
Mawalid et al. [21] | EEG | Naïve Bayes, KNN | Binary | KNN 83.3% Naïve bayes 88.9% | Detection |
Khoirunnisaa et al. [20] | EEG | SVM-RBF, KNN, LDA | Binary | SVM-RBF 83.3% KNN 83.0% LDA 100% | Detection |
Dennison et al. [25] | EEG, ECG, EOG, blink rate, breathing rate, EGG, postural sway, head movement | LDA, KNN, Naive Bayes, decision tree, ADABoostM2, and bagged decision trees | Multiclass | Unimodal Feature Bag classifier: EEG: 93.80% Posture: 83.48% Breathing rate: 81.32% HMD sensors: 78.40% Avatar movement: 74.40% ECG: 68.44% EOG: 61.84% EGG: 48.52% Multimodal feature fusion: Bag: 95% KNN: 93% ADABoost: 92% | Detection |
Wang et al. [34] | Postural sway | LSTM | SSQ score estimation | Pearson correlation coefficient r = 0.89 between SSQ score and loss (measure of difference in postural sway between pre and post VR exposure) | Detection |
Garcia-Agundez et al. [28] | ECG, EOG, blink rate, breathing rate, GSR | Fine Gaussian SVM, linear SVM, KNN | Binary and Multiclass | Binary: fine Gaussian SVM: no cs: 57.6%, minor: 74.2%, severe: 81.8% Ternary: KNN: 58% | Detection |
Jeong et al. [22] | EEG | DNN, CNN | Binary cutoff | DNN 98.02% CNN 98.82% | Detection |
Li et al. [35] | EEG, postural sway, head body movement | KNN, LR, RF, MLP in a voting classifier | Multiclass | single subject binary classification: 91.1% multiple subject binary classification: 76.3% 3 class classification: 86.7% Severe, 50.4% moderate, 79.1% mild, 68.9% average accuracy | Detection |
Kim et al. [42] | EEG | CNN, LSTM, RNN | Multiclass | LSTM with EEG 87.13% ± 1.51 Combined LSTM EEG + CNN-RNN visual features: 89.16% ± 1.87% visual predictor alone: 79.03 ± 1.24% | Detection |
Liao et al. [27] | EEG | LSTM, SVM, MLP, CNN | Binary cutoff custom sickness index | 1 min: 83.94%, 5-min 83.33%, 10 min 83.92% 82.83% for RNN-LSTM, CNN at 73.13%. MLP at 71.31% and LibSVM at 62.58% | Prediction |
Li et al. [23] | EEG | KNN, polynomial-SVM, RBF-SVM | Binary | Single subject binary classification: polynomial-SVM 92.83%, KNN 90.97%, RBF-SVM 90.74% Multiple subject classification: 79.25%, 77.5%, 73.84%, respectively | Detection |
Lee and Alamaniotis [43] | EEG | DESOM with auto encoder for clustering, KNN | Binary | DESOM Purity index 96.87% | Prediction |
Islam et al. [30] | ECG, breathing rate, GSR | LSTM regression analysis | Multiclass | MAE 8.7% | Prediction |
Islam et al. [31] | ECG, breathing rate, GSR | CNN-LSTM | Multiclass based on sickness score estimation | Detection: 97.44%, Prediction: 87.38% | Detection and Prediction |
Martin et al. [33] | BVP, EDA | SVM, GB, RF, LR | Sickness rating estimation, Binary and Multiclass | Model trained on all participants: LR: R2 0.75 RF: binary 91.7%, multiclass 86.2% One model for each participant: LR: R2 0.40 RF: binary 89%, multiclass 85.9% | Prediction |
Recenti et al. [24] | EEG, EMG, heart rate | RF, GB tree, SVM, KNN, MLP | Binary | RF: IPV 75.9%, INM, 79.5%, IMS 74.7% | Detection |
Oh and Kim [32] | BVP, respiratory signal | DELM with SVM, KNN, RF, ADAboost stacked into CNN | Multiclass | SVM: 94.23%, KNN: 92.44%, RF: 93.20%, ADABoost: 90.33%, DELM: 96.48% | Detection |