DATA REDUCTION OR DATA FUSION IN BISOGINAL PROCESSING?

When subjects are monitored over long time spans and when several biosignals are derived a large amount of data has to be processed. In consequence, the number of features which has to be extracted is mostly very restricted in order to avoid the so-called "curse of high dimensionality". Donoho (Donoho, 2000) stated that this applies only if algorithms perform local in order to search systematically for general discriminant functions in a high-dimensional space. If they take into account a concept for regularization between locality and globality "blessings of high dimensionality" are to be expected. The aim of the present study is to examine this on a particular real world data set. Different biosignals were recorded during simulated overnight driving in order to detect driver's microsleep events (MSE). It is investigated if data fusion of different signals reduces detection errors or if data reduction is beneficial. This was realized for nine electroencephalography, two electrooculography, and for six eyetracking signals. Features were extracted of all signals and were processed during a training process by computational intelligence methods in order to find a discriminant function which separates MSE and Non-MSE. The true detection error of MSE was estimated based on cross-validation. Results indicate that fusion of all signals and all features is most beneficial. Feature reduction was of limited success and was slightly beneficial if Power Spectral Densities were averaged in many narrow spectral bands. In conclusion, the processing of several biosignals and the fusion of many features by computational intelligence methods has the potential to establish a reference standard (gold standard) for the detection of extreme fatigue and of dangerous microsleep events which is needed for upcoming Fatigue Monitoring Technologies.