A comparison of methods used to detect changes in neuronal discharge patterns

https://doi.org/10.1016/S0165-0270(97)00099-XGet rights and content

Abstract

The discharge pattern of two thalamic neurones was recorded from a conscious monkey performing voluntary movements about the wrist joint. The neuronal discharge was displayed as a raster centred on movement of the wrist. The discharge patterns of both neurones was very strongly correlated with movement. Three experienced researchers were asked to examine the data and to classify every part of each trial as background discharge, `on' (increased firing rate) or `off' (decreased or zero firing rate) and to mark the times that neuronal discharge changed state. A `standard output' was made from these classifications. A back-propagation artificial Neural Network (the Network) was used to model the standard output and cumulative sums (CUSUMs) and maximum likelihood was then performed on the data and compared with the Network. There was a high correlation between the output of each observer (r>0.61) and the standard output and between the Network and the standard output (r>0.99). However the correlation between standard output and CUSUMs (r=0.06) and standard output and maximum likelihood (r=0.36) was much lower. The Network could be trained with as few as 12 trials, indicating a high degree of constancy in the methods employed by the observers. The Network was also highly efficient at detecting changes in state of neuronal activity (r>0.99). In summary, when used on single trial data, visual inspection is a reliable method for detecting timing of change neuronal discharge and is superior to CUSUM and maximum likelihood. As well it is capable of detecting neuronal discharge state: that is whether firing rate is increased, normal or decreased. Neural Networks promise to be a useful method of confirming the consistency of visual inspection as a means of detecting changes in neuronal discharge pattern.

Introduction

In many neurophysiological studies, there is a need to detect the time of changes in neuronal discharge rates. A number of techniques are available but none is satisfactory. The cumulative sum (CUSUM) technique (Ellaway, 1978) is the most commonly employed method. In our hands (Butler et al., 1992a; Butler et al., 1992b and Forlano et al., 1993), and from anecdotal discussion with workers in other laboratories, this method has the following disadvantages.

The CUSUM technique was originally applied to the analyses of spike trains as a transformation of the peri-stimulus time histogram, and this remains a common use. However, if the histogram is derived from a small number of trials, or even single trials, as is usually the case when recording from `chronic animal' preparations, particular problems arise. For CUSUM to be proficient at detecting a change, there must be only small variation in the discharge interval in the control period. With single trial data from chronic recordings, CUSUM has a high incidence of failing to detect changes that are evident on inspection and requires constant manipulation of various parameters to produce an output that compares with a reasonable expectation based on visual inspection.

Finally, there is an obligatory delay between the time when the changes in neuronal discharge rate actually occurs and the detection of that change by the cusum. This is because the CUSUM must exceed some `window' (e.g. 3 S.D. from the mean discharge rate) before a change is detected. Many laboratories employ a `walk back' to the point that 1 S.D. was crossed. However the latency between the time of change and detection of that change will vary depending on the rate of rise of the CUSUM.

In response to these difficulties we tested our perception that visual inspection was more reliable than CUSUM (Butler et al., 1992a; Butler et al., 1992b). We found each researcher to be consistent in assessment of a range of data and for there to be consistency between researchers. Visual inspection has the advantage that it can detect not only latency of onset of change in state, but also offset, decrease in discharge and increase in discharge, all of which are beyond the capabilities of the CUSUM. However, visual inspection is open to biases, is difficult to quantify and is often discouraged by journal editors.

The aim of this paper was to test whether visual inspection was more reliable than CUSUM and the maximum likelihood method, whether visual inspection was consistent over a range of data and whether a Neural Network could model visual inspection.

Section snippets

Data recording

Activity of motor thalamic neurones (VPLo, area X, VLc, VLps of Olszewski, 1952) was recorded from a monkey (Macaca nemestrina) performing skilled wrist movements in response to a visual cue. This activity was recorded as part of a separate study (Butler et al., 1996). Two neurones whose discharge patterns were clearly related to movement were chosen for this study.

For the purpose of this paper the following description of the recording method follows. Animals were trained to sit quietly with

Detection of latency of change in neuronal discharge rate

For each of the four methods described above, a correlation coefficient (Pearson's r) was calculated between the onset latency obtained with each method and the `standard output'. For the cumulative sum and maximum likelihood, a correlation coefficient with visual inspection was also calculated in order to determine whether results obtained with these methods were better correlated with visual inspection by each one of the observers than with the standard output.

It was found that some analysis

Discussion

This study demonstrates that human observers can reliably detect changes in neuronal discharge and that an artificial Neural Network can reliably model their selection criteria. On the other hand, cumulative sums and maximum likelihood proved to be far less reliable than visual inspection as judged by the high level of agreement between the three observers.

Conclusion

The results from this study suggest that visual inspection is a reliable a method for detecting changes in neuronal discharge rate. It is superior to cumulative sums and maximum likelihood for this purpose. Artificial Neural Networks can reliably model visual inspection. Furthermore they can be successfully applied to the classification of neuronal discharge into discrete states and are worthy of further investigation to apply them to a wider range of neuronal discharge patterns, a larger

References (9)

There are more references available in the full text version of this article.

Cited by (18)

  • Use of cumulative sum (CUSUM) test for detecting abrupt changes in the process dynamics

    2007, International Journal of Reliability, Quality and Safety Engineering
View all citing articles on Scopus
View full text