Key Feature Extraction for Fatigue Identification using Random Forests

Electroencephalogram (EEG) might be the most predictive and reliable physiological indicator of mental fatigue. However, the extraction of key features from massive EEG data for mental fatigue identification remains a challenge. The objective of this study is to identify the key EEG features in relationship to mental fatigue, from a broad pool of EEG features generated by quantitative EEG (qEEG) techniques, using Random Forests (RF), which is a recently developed machine learning algorithm. The method is applied to key EEG feature extraction for 5-level mental fatigue identification using the five subjects' EEG data recorded in 25-hour fatigue experiments. RF produces significant feature reduction with little compromise of the classification performance. The identified key EEG features also indicate that electrode locations in frontal and occipital regions of the brain are most important for adequate representation of the deactivation of functional lobes of the brain, which is consistent with the anatomical areas known to be involved in mental fatigue. It is also interesting to discover that the four frequency bands are all important for the mental fatigue identification.