|
Article Excerpt
INTRODUCTION
In human-computer interaction, many tasks are performed using complicated cognitive information processing. Such situations cause mental workload or stress that has not been observed as a result of performing traditional physical tasks. As a result of workers' reduced agility caused by excessive workload, productivity decreases and human error frequently occur. In particular, mental workload induced by overloading the capacity of working memory would be a limiting factor in the early stage of acquiring computer procedural skills. An effective method of monitoring mental workload would therefore be useful in the evaluation of alternative interface designs.
Heart rate variability (HRV) is frequently used as a measure for evaluating mental workload (Aasman & Mulder, 1987; Akselrod et al., 1981; Bartoli, Baselli, & Cerutti, 1985; Baselli & Cerutti, 1985; Baselli et al., 1987; Cerutti, Fortis, Libeart, Baselli, & Civardi, 1988; Luczak & Laurig, 1973; Murata, 1994; Tsuji & Mori, 1994). Cerutti et al. (1988) applied an autoregressive model to analyze the fluctuation of R-R intervals and showed that the difference in mental activity between a resting condition and that required to perform a mental arithmetic task could be evaluated on the basis of a low-frequency component at around 0.1 cycles/beat and a high-frequency component corresponding to the respiration-frequency band. Murata (1994) also indicated that mental workload caused by demanding tasks such as arithmetic could be evaluated using an autoregressive power spectrum if the respiration interval was carefully controlled. HRV is a measure that indicates the activity of the autonomic nervous system. Physiological measures such as an electroencephalogram (EEG) might also be promising, especially for the evaluation of mental workload induced by activities such as memory tasks.
EEG measures have been shown to be highly sensitive to variations in task difficulty (Gevins et al., 1998; Gevins, Smith, McEvoy, & Yu, 1997; Gevins, Zeitlin, Doyle, Schafer, & Callaway, 1979; Gevins, Zeitlin, Doyle, Yingling, et al., 1979; Gundel & Wilson, 1992; Humphrey & Kramer, 1994; Wilson & Fisher, 1995). In these traditional analyses, the power spectral density of EEG signals is calculated using fast Fourier transform (FFT) to examine the change of frequency characteristics. Such an approach allows one to understand how the ratio of a specific frequency band such as an alpha band changes with the accumulation of mental fatigue (Okogbaa, Shell, & Filipusic, 1994) or when the mental work level changes (Gevins et al., 1998). The power spectrum represents frequency characteristics during a predetermined interval. If EEG is measured at different time epochs, one can roughly know how the frequency characteristics change for different time epochs. Unlike time series data, however, the time characteristics during the interval cannot be known.
It has been shown that the P300 component of the event-related potential (ERP) can be an effective measure of mental or cognitive workload (Johnson & Donchin, 1980; Kramer, Wickens, & Donchin, 1985; Ullsperger, Metz, & Gille, 1988). The P300 amplitude increases with increased cognitive work intensity. However, it takes a great deal of time to obtain an ERP waveform 2based on each EEG recording. Cognitive information processing is conducted in the central nervous system in a very short period. Such a time-consuming measurement using ERP (P300) has the potential to miss the change in cognitive information processing according to the difference of mental work loading.
EEG signals are, usually, nonstationary and change abruptly in a short period, which leads to the change of frequency characteristics. It is impossible to detect such a change of frequency characteristics by means of traditional approaches using power spectral analysis of EEG or the P300 component of ERR Theoretically, signal processing of EEG signals by FFT assumes that the signals are stationary. Biological signals are composed of phenomena that change every moment, often abruptly. A time series that does not contain such changes can be regarded as stationary. Actually, biological signals such as EEG and R-R intervals are not stationary. The wavelet analysis or short-time Fourier transform (STFT), which can investigate the time-frequency characteristics of biological signals, is expected to extract more useful information that cannot be extracted with traditional approaches.
The STFT is a useful technique in that it allows one to depict nonstationary signals as stationary ones by using a window function. The time-frequency resolution of STFT is fixed over an entire time-frequency plane once a window is chosen. In such nonstationary signals, the change in stress or workload factors results in an abrupt change in biological signals. The STFT does not properly treat such changes in biological signals. Better resolution in time at higher frequencies (rapid changes in time) is needed. At low frequencies, better resolution in frequencies is required. To overcome these problems, the wavelet transform (Daubechies, 1988, 1990) is promising because it changes the window size adaptively and uses short windows at high frequencies and long windows at low frequencies. Consequently, it is expected that the wavelet transform is applicable to nonstationary signals.
In spite of this, biological signals such as R-R intervals and EEG are analyzed using an arbitrary analysis interval that is assumed to be stationary. It is reasonable to assume that the mental state or mental workload is nonstationary and changes every moment. Global analysis using a traditional signal processing technique such as FFT and local analysis that investigates the time-frequency characteristics of biological signals are both necessary in the evaluation of mental workload. Therefore, it is promising...
|
