Automatic relevance determination (ARD) was applied to two-channel EOG recordings for microsleep event (MSE) recognition. 10 s immediately before MSE and also before counterexamples of fatigued, but attentive driving were analysed. Two type of signal features were extracted: the maximum cross correlation (MaxCC) and logarithmic power spectral densities (PSD) averaged in spectral bands of 0.5 Hz width ranging between 0 and 8 Hz. Generalised learning vector quantisation (GRLVQ) was used as ARD method to show the potential of feature reduction. This is compared to support-vector machines (SVM), in which the feature reduction plays a much smaller role. Cross validation yielded mean normalised relevancies of PSD features in the range of 1.6 -4.9 % and 1.9 -10.4 % for horizontal and vertical EOG, respectively. MaxCC relevancies were 0.002 -0.006 % and 0.002 -0.06 %, respectively. This shows that PSD features of vertical EOG are indispensable, whereas MaxCC can be neglected. Mean classification accuracies were estimated at 86.6  1.3 % and 92.3  0.2 % for GRLVQ and SVM, respectively. GRLVQ permits objective feature reduction by inclusion of all processing stages, but is not as accurate as SVM.

Automatic relevance determination (ARD) was applied to two-channel EOG recordings for microsleep event (MSE) recognition. 10 s immediately before MSE and also before counterexamples of fatigued, but attentive driving were analysed. Two type of signal features were extracted: the maximum cross correlation (MaxCC) and logarithmic power spectral densities (PSD) averaged in spectral bands of 0.5 Hz width ranging between 0 and 8 Hz. Generalised learning vector quantisation (GRLVQ) was used as ARD method to show the potential of feature reduction. This is compared to support-vector machines (SVM), in which the feature reduction plays a much smaller role. Cross validation yielded mean normalised relevancies of PSD features in the range of 1.6 -4.9 % and 1.9 -10.4 % for horizontal and vertical EOG, respectively. MaxCC relevancies were 0.002 -0.006 % and 0.002 -0.06 %, respectively. This shows that PSD features of vertical EOG are indispensable, whereas MaxCC can be neglected. Mean classification accuracies were estimated at 86.6  1.3 % and 92.3  0.2 % for GRLVQ and SVM, respectively. GRLVQ permits objective feature reduction by inclusion of all processing stages, but is not as accurate as SVM.