EEG-based Drowsiness Detection using Support Vector Machine

Every year 100,000 car crashes occur as a direct result of driver fatigue resulting in 1,550 deaths, 71,000 injuries, and $12.5 billion in losses as reported by the National Highway Traffic Safety Administration (NHTSA). Another study done by the National Sleep Foundation (NSF) found that drowsiness causes up to 1.9 million crashes a year, with 54% of all drivers having driven while drowsy at least once in the past year, and 28% doing it monthly. It is obvious that drowsiness is a major problem not only from a safety standpoint, but also from a financial one as well. Numerous studies exist in the neuroscience and engineering disciplines that attempt to determine the moment when a person starts to lose focus and drift into sleep. This state change can be detected by various behavioral and physiological measures such as yawning, eye blinking, electroencephalogram (EEG), and electrocardiogram (ECG). There is no general consensus on which of these methods is the most effective and some studies even adopt a hybrid approach in order to maximize the accuracy. The brain?s role at the center of the nervous system means it is actively aware of the current cognitive state of an individual, and by measuring the synchronization of neurons fired by various regions of the brain we can acquire a glimpse into this state. Simple events such as eye closures can be easily distinguished from the resulting EEG waveform while more complex mechanisms such as heart abnormalities and sleep stage transitions require a more thorough understanding of the topic in order to develop effective feature extraction and selection methods.

The work presented in this thesis leverages the brain activity collected in the form of electrical currents, known as EEG, with Support Vector Machines (SVM) to create a classifier which incorporates machine learning that is capable of detecting drowsiness with high accuracy. Traditional drowsiness detection methods utilize the generally defined broadband definitions of the EEG frequency bands in their detection and feature selection algorithms. However, due to the inherently complex nature of EEG signals, particularly with regards to interpersonal differences such as age and health, these drowsiness detection methods for the most part are not able to achieve very good results in terms of accuracy and precision. In addition, recently discovered phenomena point toward unrelated behavior in sub-bands within certain EEG bands, in particular the alpha band which is composed of three sub-bands (one lower and two upper). Contradictory behavior in the lower and first upper alpha band has been observed depending upon the nature and condition of drowsiness. Coupled with the lack of general consensus on the specific sleep stage which encompasses the drowsy state, drowsiness detection becomes a multi-facetted problem which includes many complexities.

Our approach takes advantage of neuroscientific evidence of the awake to sleep transition as a guide for feature selection. To facilitate this process, sub-banding was adopted in order to provide high resolution analysis of the EEG signal as well as flexibility in the selection of frequency ranges for use in the analysis. In addition, techniques to account for interpersonal variability were developed. By selecting a compact set of features that offer the greatest differentiability between the awake and drowsy states, we trained and tested a SVM classifier which provided high performance with up to 97.94% overall accuracy in classifying between awake and sleep stage 1 during testing.