Attention or vigilance monitoring has been studied extensively over the years, in particular in connection with tasks which require a certain degree of attention or vigilance in order to preserve safety. For example, in repetitive but safety-critical tasks such as driving a train, car or other vehicle, or in complex tasks such as air- traffic control, it is important that the subject remains vigilant in order to conduct the task safely. However, it has been difficult to find a way of monitoring vigilance effectively and objectively, particularly in a way which is not intrusive to the operator.

A state of alertness can, in essence, be regarded as a state which is part of a continuum from a completely alert state to a completely (deep) sleep state. Thus vigilance monitors based on the extension of sleep analysis, particularly of electroencephalograms (EEG), have been proposed, which analyse the spectral content of the EEG to track the state of alertness/drowsiness of the subject. One example of such a monitor (disclosed in, for example EP-A-0773504) uses an artificial neural network to analyse the EEG, the neural network being trained on representative samples of alert EEG and drowsy EEG as assessed by a vigilance expert. However, a problem with this is that the ability of a subject to perform tasks requiring a certain level of attention or vigilance, is not necessarily correlated well with the subject's EEG characteristics as assessed by the expert.

The paper"Estimating Alertness from the EEG Power Spectrum"by Jung et al. (IEEE Transactions On Biomedical Engineering. Volume 44, No. 1, January 1997) describes a method of estimating alertness from an EEG power spectrum in which subjects are required to respond to audible and visual stimuli while having their EEG monitored. The number of times that the subject misses an audible stimulus within each minute of the test is taken as a measure of alertness, and its correlation with the EEG was examined. However, both positive and negative

correlation with the EEG were found, and the results from many subjects had to be combined to obtain enough data, making the results unsuitable as the basis for a satisfactory vigilance monitor. Furthermore, the measurement of the number of misses of stimuli per minute is a poor way of measuring alertness, because alertness can vary significantly on a timescale smaller than one minute, and in a more subtle way than is detected by characterising responses as coarsely as the number of misses or hits.

The present invention provides an improved performance test which gives a continuous performance measure. This test can be used as the basis for an improved vigilance monitor.

Thus one aspect of the invention provides an apparatus for measuring the psychomotor performance of a subject comprising a continual stimulus for the subject, a sensor for detecting a discrete voluntary motor response of the subject to the stimulus, a timer for measuring the timing of the response, a processor for calculating from the variations in the timing a continuous performance measure, a biosignal sensor for measuring a biosignal of the subject and a trainable model relating the biosignal to the continuous performance measure.

The continual stimulus may a be visual stimulus such as a light which is repeatedly illuminated, for instance at regular intervals for a constant period of time.

For example, the light may be a dimly illuminated light emitting diode (LED), and in one version of the test it may be illuminated for one second every three seconds, though clearly other illumination patterns, or indeed visual stimuli, are usable. The sensor may be a movement detector, which could be a proximity detector or a button operating a switch. The timer may time the response in relation to the stimulus.

Where the stimulus is regular, the timer may time the intervals between successive responses or the timing relative to the stimulus itself. The continuous performance measure may be based on the variation in the intervals. If the subject is completely alert then the intervals will be consistent (as the subject's response time will be approximately the same each time). If, however, the subject becomes less alert, there will be an increase in the variability of the timing of responses, resulting from a

delayed reaction to the stimulus.

The continuous performance measure may conveniently be a weighted average of the intervals calculated over a moving time window.

Preferably the apparatus further comprises an electroencephalographic monitor for acquiring as a biosignal the electroencephalogram (EEG) of the subject.

The processor is then adapted to analyse the EEG and determine the relation between changes in the EEG and changes in the psychomotor performance. Preferably the processor analyses the frequency content of the EEG, for instance using a parametric model such as an auto-regressive model or Fast Fourier Transform.

Preferably a visual stimulus is the only stimulus to the subject, and the apparatus may include an audio isolator, such as a white noise generator and/or appropriate sound insulation of the test environment.

The trainable model may be an artificial neural network, described in more detail below.

The use of the timing of the response to give the continuous performance measure allows the construction of an analogue scale from a series of discrete events.

In essence the timings form a scale on the basis of which the performance measure is established. Of course the times may be measured digitally, but they are measured at a sufficiently high resolution as to form a quasi-continuous scale.

Another aspect of the invention provides a method of determining the relation between a biosignal of a subject and the vigilance level of the subject, comprising the steps of monitoring the biosignal of the subject while the subject performs a psychomotor performance test using a stimulus, detecting a discrete voluntary motor response of the subject to the stimulus, measuring the timing of the response and calculating a continuous performance measure from the variations in timing.

The performance test may be that in accordance with the first aspect of the invention. The biosignal and performance measure may be used to set up an analyser, eg to train an artificial neural network, to determine the level of vigilance of the subject from solely the biosignal.

Thus another aspect of the invention provides a vigilance monitor for

monitoring the level of vigilance of a subject comprising a biosignal monitor for acquiring a biosignal of the subject, and a processor for analysing changes in the acquired biosignal to produce a measure of the vigilance of the subject, wherein the processor analyses the biosignal on the basis of the relation between the biosignal and the vigilance level previously obtained by the method above.

The processor may comprise a trained artificial neural network, the neural network having been trained on a data set comprising the frequency content of a biosignal acquired during a psychomotor performance test as described above.

As with the first aspect the biosignal may be an EEG. The relation between the biosignal and the vigilance level may be subject specific. It may be determined by performing the performance test on an individual and then incorporating that relationship into a vigilance monitor adapted to that individual. For instance, where the monitor uses an artificial neural network it may be trained on data produced by testing the individual concerned. One way of doing this would be to test subjects over a two-day period, one day being after normal sleep and the second day after a period of sleep deprivation. The data from the performance tests is used to configure the vigilance monitor to that person's own characteristics. The vigilance monitor can then be used by them (their vigilance being assessed only by reference to their biosignal-e. g. EEG) without the need for further performance testing.

The invention will be further described by way of example with reference to the accompanying drawings in which:- Figure 1 is a flow diagram schematically illustrating the performance test according to one embodiment of the present invention; Figure 2A illustrates apparatus for conducting the performance test illustrated in Figure 1; Figure 2B illustrates the training of an artificial neural network for analysing EEG data; Figure 2C illustrates an embodiment of a vigilance monitor using the trained artificial neural network of Figure 2B; Figure 3 is a flow diagram schematically illustrating the analysis of the EEG

and performance data in an embodiment of the invention; Figure 4 is a flow diagram schematically illustrating the derivation of the performance measure in Figure 3; and Figure 5 is a graph illustrating the results from testing a specific subject.

One embodiment of a performance measure in accordance with the invention and of a vigilance monitor based on it will now be described.

The performance test used in this embodiment is in some ways similar to a behavioural test used for testing resistance to sleep. This is the Oxford Sleep Resistance Test (OSLER test described in J. Sleep Res. (1997) 6,142-145 by L S <BR> <BR> Bennett et al. ). The test involves four test sessions during a day, each of 40 minutes duration. In the example reported on below the tests were conducted at 0900,1100, 1300 and 1500 hours. Subjects are required to abstain from alcohol for 24 hours and from coffee for 12 hours before the day of study, and to abstain from sleeping during the day of study. As indicated schematically in Figure 1, in each 40 minute test, in step S1 the subject is placed semi-recumbent in a darkened room and asked to respond to visual stimuli. In this example the visual stimuli are one second long dim illuminations of an LED every three seconds as indicated in step S2. The subject responds by pressing a button indicated as step S3. The button presses and EEG are recorded using a PC and if the test time has not expired (S4) the illumination is repeated. If seven flashes are missed consecutively (S5) the test is stopped as the subject is deemed to be asleep. The test ends at S6. The equipment required is illustrated schematically in Figure 2A. The subject may wear headphones 2 connected to a white noise generator 4 which provides audio isolation for the subject and avoids any stimulus other than a visual one. The subjects are requested to respond to the illumination of a dimly lit LED 6 by pressing a button 8 on a hand piece. The illumination of the LED is controlled by the computer system 10. The computer system 10 records the timing of the button presses and, optionally, the timing of the illumination of the LED 6. Further, a single channel of EEG is recorded using standard EEG monitor 11 and contact electrodes 12. The EEG may

be recorded from either the central, frontal, or mastoid sites. The signals from button 8 may be sampled at the same sampling frequency as the EEG, for instance 128 Hertz or 256 Hertz, though a lower sampling rate is acceptable for the button timings.

The process flow is illustrated in Figure 3. In step 30 the EEG and button press timings are recorded for one or more subjects using the apparatus illustrated in Figure 2A. Both the EEG and the button press timings are stored for later analysis.

Following all four tests, the EEG recording is parameterised, in this embodiment, by fitting an autoregressive model to each Is epoch of EEG data as illustrated in step 34. The autoregressive parameters describe the frequency content of the epoch under consideration. Alternatively, other frequency content descriptions such as the Fast Fourier Transform (FFT) could be used. If the test ran for the full 40 minutes, there would therefore be 2400 resulting epochs of EEG for that test and a corresponding 2400 vectors of autoregressive parameters. Artefactual epochs in the EEG are identified by monitoring for electrical amplitudes which are unlikely to be caused by electrical potentials within the brain. These epochs are excluded from further analysis in step 32.

The analysis of the button press data of step 36 is illustrated in more detail in the process flow of Figure 4. During the test, in step 40, the interval between consecutive button presses is recorded. It will be appreciated that with a perfectly vigilant subject, this interval will be close to three seconds, corresponding to the interval between illuminations of the LED. However, as the subject becomes less vigilant, this time will increase. In order to calculate a continuous performance measure, in step 41 the squares of the difference between the intervals and three seconds is calculated. These values are then filtered in steps 42 and 43 by using a four point moving average filter, first forwards through the time series and then backwards through the time series. This filter involves adding together the values for four successive points and dividing the result by four (a filter using a coefficient other than 1/4 may be used). Conducting this in the two steps 42 and 43, in the forward and reverse directions, results in a filtered value for each button press which

is in phase with the original values. The square root (step 44) of these filtered values gives as an output the continuous performance measure in step 45.

Figure 5 illustrates just under ten minutes of data from a subject in a specific test. The black dots are plots of the time between button presses, and it can be seen that at the beginning of the test most button presses occur at around three second intervals, showing that the subject is fairly vigilant. Occasional misses result in a longer time between button presses. As the test proceeds, though, the time between button presses becomes much more variable, indicating a decrease in vigilance. The continuous performance measure is indicated by line 50 on the graph of Figure 5.

This continuous performance measure provides sub-minute scale information on the vigilance of the subject, which is important in accurately and correctly assessing vigilance and necessary to enable the optimisation of vigilance monitoring. By way of comparison the line 52 plots the number of missed hits per minute. The advantage of the continuous performance measure indicated by line 50 is that regions where performance is poor, but where very few illuminations were actually missed, are correctly identified as representing poor performance which is not true of the number of missed hits per minute measure.

Returning to the process flow of Figure 3, each of the epochs of the EEG is associated with a particular value of the performance measure during that epoch. In this embodiment the range of performance measures is divided into 30 equal divisions and the epochs of the EEG are assigned to one of those divisions according to their performance measure in step 36.

In order to provide a vigilance monitor based on this performance test, the EEG data 14 and the performance data 15 are used to train an artificial neural <BR> <BR> network 16 (such as a multi-layer perceptron (MLP) ) as illustrated in Figure 2B. The trained artificial neural network will be applicable only to the subject under consideration as the precise nature of changes in the EEG which reflect loss of vigilance are subject-specific and the effectiveness of the trained artificial neural network can only be fully established using the data made available by the performance testing of the particular subject. Techniques for training artificial neural

networks are well-known and will not be described in detail here. Details can be <BR> <BR> found in, for instance, "Neural Networks for Pattern Recognition"by Christopher M.

Bishop (Oxford University Press).

The trained neural network 16 is illustrated in use in vigilance monitor 100 in Figure 2C. EEG data obtained by EEG monitor 11 via contact electrodes 12 is fed to the trained neural network which produces a vigilance measure which is displayed and/or plotted on vigilance indicator 80 and which is also used to trigger an alarm if the level of vigilance of the subject falls below a predetermined limit.

It should be appreciated that optimally the artificial neural network is trained with data from the specific subject whose vigilance is to be monitored by the monitor 100. This results in the optimal correlation between the level of vigilance estimated and the EEG of the subject. However, the trend of increasing power at low frequencies in the EEG corresponding to a lower level of vigilance is common in the population, and so the vigilance monitor 100 is still effective if monitoring a subject different from the one whose data was used to train the artificial neural network, or data from many subjects may be used to train a generic artificial neural network.

It should also be appreciated that while this embodiment uses an artificial neural network to analyse the EEG data, alternative embodiments can use different techniques once the correlation between the EEG and the level of vigilance has been established. Thus, for instance, the vigilance monitor could monitor the relative power in certain specific frequency bands and calculate the level of vigilance from that.