Prediction of microsleeps from EEG using Bayesian approaches.
Thesis DisciplineElectrical and Electronic Engineering
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
Microsleeps are brief unintentional episodes of sleep-related loss of consciousness up to 15 s during an active task, such as driving. Such episodes of unresponsiveness are of particularly high importance in people who perform high-risk occupations requiring extended and unimpaired visuomotor performance, such as truck drivers, train drivers, pilots, and air-traffic controllers, where microsleeps can, and do, result in catastrophic accidents and fatalities. Therefore, prediction and even early detection of microsleeps has the potential to prevent sleep-related accidents and save lives.
The aim of this study was to investigate Bayesian methods in the prediction of microsleeps from the EEG. Bayesian methods were also investigated as a measure of improving performance in microsleep detection. This study used data from a previous study, ‘Study A’, in which 15 healthy non-sleep-deprived participants performed two 1-h sessions of 1-D continuous tracking task (CTT). Participants who experienced at least one definite microsleep during the two CTT sessions, i.e., 8 subjects, were included for the analyses. The original gold-standard was formed by integrating two independent measures: the video-rating from a human expert and tracking flat-spots.
In this study, we first refined the gold-standard to improve its accuracy and minimize potential false labels. A microsleep was defined as a non-tracking episode in conjunction with a deep-drowsy or lapse video rating, whereas a responsive episode was defined as a satisfactory tracking performance for at least of 5 s irrespective of video ratings. The remainder of the gold-standard was labelled uncertain and pruned out. We then proposed four Bayesian models for feature reduction. The first Bayesian model was variational Bayesian robust factor analysis (VBRFA), which is an extension of variational Bayesian FA (VBFA) that finds a robust latent space. This model was then extended to variational Baysian multi-subject RFA (VBMSRFA), which assumes independent mean and noise terms for individual subjects. Variational Bayesian hierarchical MSRFA (VBHMSRFA) finds a group level loading matrix and allows individual subjects to have slightly different loading matrices. Finally, variational Bayesian hierarchical multi-subject robust joint matrix factorization (VBHMSRJMF) incorporates the information of class labels while finding a less subject-variable latent space. All of these proposed methods used variational inference to approximate the posterior probabilities. Moreover, automatic relevance determination (ARD) motivated prior distributions were used in all models to automatically find the optimum number of latent variables.
Various features of EEG data were extracted and investigated, namely power spectral features (PSF) (192 features), power spectral features using individual alpha frequency (PSF-IAF) (192 features), multiple domain features (MDF) (176 features), wavelet mean squared features (WMSF) (80 features), wavelet log mean squared features (WLMSF) (80 features), and wavelet energy percentage features (WEPF) (80 features). These features were extracted from 2-, 5-, 10-s window lengths. In addition, four classifiers, i.e., linear discriminant analysis (LDA), linear support vector machine (SVM), tree augmented naïve Bayes (TAN), and variational Bayesian logistic regression (VBLR), were investigated to discriminate microsleeps. We found that aggregating features of the three window-lengths performed superior to individual single-window features. The proposed Bayesian feature reduction methods were applied to each feature set and the meta-features were extracted. These meta-features were then fed to a classifier to perform detection and prediction of microsleeps.
The best microsleep state detection performance was phi correlation coefficient (phi) – a performance measure for imbalanced datasets to quantify the correlation between actual and predicted labels – phi = 0.47 with an LDA and VBMSRFA meta-features of WLMSF (AUC-ROC = 0.95; AUC-PR = 0.49; GM = 0.83; Sn = 0.74; Pr = 0.38). On the other hand, the highest performance of the original features without any feature reduction/selection method, i.e., baseline performance, was ϕ = 0.40, achieved with a linear SVM and PSF (AUC-ROC = 0.95; AUC-PR = 0.49; GM = 0.79; Sn = 0.74; Pr = 0.38).
With a prediction time of τ = 1 s for microsleep state prediction, the highest performance was ϕ = 0.44 which was achieved with an LDA classifier and VBMSRFA meta-features of WLMSF (AUC-ROC = 0.94; AUC-PR = 0.44; GM = 0.80; Sn = 0.69; Pr = 0.36). In contrast, the highest baseline performance was ϕ = 0.35 with a linear SVM and MDF (AUC-ROC = 0.92; AUC-PR = 0.42; GM = 0.76; Sn = 0.70; Pr = 0.35). Overall, VBMSRFA meta-features of WLMSF led to the highest performance for both detection and prediction of microsleep states.
The highest performance in terms of AUC-ROC and AUC-PR for microsleep onset detection was achieved with a linear SVM and PSF (AUC-ROC = 0.91; AUC-PR = 0.09; ϕ = 0.08; GM = 0.71; Sn = 0.79; Pr = 0.03). Notwithstanding, our proposed Bayesian methods achieved slightly higher GM and phi values. The highest phi of 0.10 (AUC-ROC = 0.89; AUC-PR = 0.05; GM = 0.78; Sn = 0.70; Pr = 0.03; Sp = 0.88) was achieved with a VBLR classifier and VBMSRFA meta-features of WLMSF. The highest GM, however, was 0.81 (AUC-ROC = 0.91; AUC-PR = 0.05; ϕ = 0.09; Sn = 0.77; Pr = 0.02; Sp = 0.85) with the same meta-features but with a linear SVM classifier.
With a prediction time of τ = 1 s, the highest values of GM and phi for microsleep onset prediction were 0.80 and 0.08, respectively, and were achieved with a linear SVM and VBHMSRFA meta-features of MDF (Sn = 0.81; Pr = 0.01; Sp = 0.80). The highest performance of the baseline with the same prediction time was GM = 0.71 and ϕ = 0.07, achieved with a linear SVM and MDF (Sn = 0.76; Pr = 0.01; Sp = 0.74). Increasing the prediction time to τ = 10 s, the highest performance for microsleep onset prediction was seen with a linear SVM and VBHMSRFA meta-features of MDF (ϕ = 0.04; GM = 0.66; Sn = 0.66; Pr = 0.01; Sp = 0.71), while the highest performance of the baseline was with a linear SVM and WMSF (ϕ = 0.04; GM = 0.54; Sn = 0.57; Pr = 0.01; Sp = 0.73).
Our findings suggests that taking inter-subject variability into account can improve accuracy of microsleep detection and prediction by reducing the variability of classification threshold between training and test subjects. However, although our results indicate that microsleeps can be better detected and predicted by incorporation of Bayesian methods, the performances are still too low for real-life applications. Further investigations are needed to find a substantially improved microsleep prediction system for real-life applications.