Automated assessment of vigilance using committees of radial basis function analysers

It is clear that whenever people perform repetitive, boring or long-term tasks, a loss in concentration can occur. These lapses in vigilance may have serious consequences under certain circumstances. Human cognitive processes are far from understood, and the processes by which individuals have lapses in alertness are many. In this paper, we consider the analysis of the human EEG during vigilance experiments as a case study in supervised data analysis. We address several issues which are found in many data analysis problems. The issue of finding informative signal parameterisations in multi-channel environments is approached using a feature selection process. Subsequent analysis uses Bayesian committees of radial basis function analysers. A comparison is made of two analysis approaches, the first based on regression and the second setting the problem as a classification task with extremal-label training. Results are presented for a representative sample of subjects. The assessment of vigilance from a small number of physiological measurements may be of importance in safety monitoring in a number of professions. We have shown that it is possible to make reasonable estimates of the state of alertness of a subject based on EEG and eye-movement information. The estimates appear to be fairly robust across subjects using appropriately-chosen features from the signals. However, in general, neither the smoothed labels (human-scored) nor the resultant estimated measures correlate well with the corresponding tracking performance measure.