Table 6.
Biological-based drowsiness detection systems.
| Ref. | Biological Parameters | Sensors | Extracted Features | Classification Method | Description | Quality Metric | Dataset |
|---|---|---|---|---|---|---|---|
| [79] | Brain activity | Bluetooth-enabled EEG headband and a commercial smartwatch | Relative EEG power ratio (power percentages) |
SVM-based posterior probabilistic model | A real-time system used an SVM-based posterior probabilistic model to detect and classify drowsiness into three levels. | Accuracy: Drowsy case: 91.92% Alert case: 91.25% Warning case: 83.78% |
Prepared their own dataset |
| [80] | Brain activity | EEG (silver surface electrode) |
IMF of the EEG signal | ANN | Detection was based on the extraction of the IMFs from the EEG signal by applying the EMD method. | Accuracy: 88.2% | Prepared their own dataset |
| [81] | EEG signals and EEG spectrogram images | EEG Sensors | Energy distribution and zero-crossing distribution of the raw EEG signals, in-depth features of the EEG spectrogram, etc. | LSTM network | EEG-based drowsiness detection method. It used pre-trained AlexNet and VGG16 models to extract in-depth features from the EEG spectrogram images. | Accuracy: 94.31% | MIT/BIH polysomnographic EEG database [82] |
| [83] | EEG | EEG Sensors | The first quartile, median, range, and energy of the Hermite coefficients | ELM decision tree, KNN, least squares SVM, ANN, and naive Bayes |
Detection was based on an adaptive Hermite decomposition for EEG signals. The Hermite functions were employed as basic functions. | Accuracy: ELM: 92.28% Sensitivity: ELM: 95.45% |
MIT/BIH polysomnographic database [82] |
| [85] | EEG | Standard wet-electrode EEG and a cap-type dry-electrodeEEG | Multi-taper power spectral density | Extreme gradient boosting classifier | A framework for detecting instantaneous drowsiness with a 2-s length of EEG signal. It was implemented on a wireless and wired EEG to show its applicability in a mobile environment. | Accuracy: Wired EEG: 78.51% Wireless EEG: 77.22%. Sensitivity: Wired EEG: 78.5%, Wireless EEG: 68.3% |
Prepared their own dataset |
| [87] | EEG | EEG sensors | F1–F9, extracted from Higuchi fractal dimension, complexity, and mobility characteristics of the original EEG signal, as well as all the EEG sub-bands |
Extra trees classifier | Employed wavelet packet transform to extract the time domain features from a single-channel EEG signal. Eleven classifiers were tested in this work. The extra trees classifier had the best results. | Accuracy, sensitivity, and precision: Dataset1: 94.45%, 95.82%, and 96.14% Dataset2: 85.3%, 79.55%, and 90.02% |
Dataset1: Fpz-Cz channel dataset [89,90] Dataset2:SVDD dataset [91] |
| [96] | EEG | EEG Sensors | Tsallis entropy, Renyi entropy, permutation entropy, log energy entropy, and Shannon entropy | Ensemble boosted tree classifier | Used AVMD to analyze and synthesize the EEG signals. By applying statistical analysis, five entropy-based features were selected. Ten classifiers were used, and the ensemble boosted tree classifier achieved the highest accuracy. | Accuracy: 97.19% Sensitivity: 97.01% Precision: 98.18% |
MIT/BIH polysomnographic dataset [82] |
| [100] | Heart rate and blood volume changes | ECG and PPG | Features obtained from Bin-RP, Cont-RP, and ReLU-RP patterns | CNN | Used wearable ECG/PPG sensors to track the different patterns in HRV signals in a simulation environment and used CNN. | Best accuracy, sensitivity, and precision: ECG: 70%, 85%, and 71% PPG: 64%, 78%, and 71% |
Prepared their own dataset |
| [101] | Heart rate | PPG | Frequency measurements (HF, LF, and HF/LF) extracted from PPG signals | Differentiating between two (HF, LF, and HF/LF) patterns | Detection is done by analyzing the changes in PPG signals frequency measurements (HF, LF, and HF/LF) that are obtained from measurements on fingers and earlobes | Accuracy: 8/9 = 88.8% | Prepared their own dataset |
| [102] | Heart rate | Wrist-worn wearable sensor and ECG sensor |
HRV and activity of the autonomic nervous systems | Random Tree, RF, KNN, SVM, Decision Stump, etc. | Detection was based on the physiological data extracted from a wrist-worn wearable sensor and ECG sensor. Multiple ML algorithms for binary classification were used | The highest accuracy was more than 92% for the KNN algorithm | Prepared their own dataset |
| [103] | HRV | ECG electrodes | MeanNN, SDNN, RMSSD, TP, NN50, LF, HF, and LF/HF | Multivariate statistical process control | Detection was based on HRV analysis. Eight HRV features were monitored to detect the changes in HRV using the multivariate statistical process control anomaly detection method. The algorithm was validated by comparing its results with EEG-based sleep scoring. | Accuracy: 92% | Prepared their own dataset |
| [104] | Respiration | Three respiratory inductive plethysmography sensors | RRV and quality of the respiratory signals | Thoracic effort-derived drowsiness | An algorithm for DDD, based on the respiratory signal variations. It combined the analysis of the RRV and the quality level of the respiratory signals to detect the changes in the driver’s alertness status. | Sensitivity: 90.3% | Prepared their own dataset |
| [107] | ECG and EMG | Disposable Ag–AgCl electrodes | Features extracted from the bispectrum of the signals H1, H2, and H3 | Linear discriminant analysis, quadratic discriminant analysis, and KNN classifiers | Detects hypovigilance caused by drowsiness and inattention using ECG and EMG signals. The gathered physiological signals from the experiments were first pre-processed. Then, multiple higher-order spectral features were extracted to be classified. | Accuracy and Sensitivity: ECG with KNN: 96.75% and 98% EMG with linear discriminant analysis: 92.31% and 96% Fused features with KNN: 97.06% |
Prepared their own dataset |
| [108] | ECG and EMG | Two pieces of conductive knit fabric | EMG peak factor and maximum of the cross-relation curve of ECG and EMG | Discriminant criterion using Mahalanobis distance | A noncontact onboard DDD system studied the EMG and ECG signals changes during driving. Feature selection was applied using the Kolmogorov–Smirnov Z test. | Accuracy: 86%. Sensitivity: 91.38% Precision: 83.45% |
Prepared their own dataset |
| [109] | ECG and EEG |
Enobio-20 channel device | EEG signals time-domain statistical descriptors, complexity measures, power spectral measures, ECG signals HR and HRV’s LF, HF, and LF/HF ratio | SVM | Combined ECG and EEG features to detect drowsiness. After the feature extraction, a paired t-test was only used to select the significant features. | Accuracy: 80.9% | Prepared their own dataset |
| [110] | EEG, EOG, ECG | EEG, ECG, and EOG electrodes | EEG features from the temporal, frontal, and occipital channels EOG features: eyeblink rate ECG feature: blood pressure and heart rate |
Linear discriminant analysis, linear SVM, kernel SVM, and KNN | The fuzzy mutual information-based wavelet packet transform method extracted the features. The features were dimensionally reduced, using spectral regression and kernel-based spectral regression methods. After that, four classifiers were applied. | Accuracy: Spectral regression: 95% Kernel spectral regression: 97% |
Prepared their own dataset |