Failure in identification of overlapping spikes from multiple neuron activity causes artificial correlations

Abstract

Recording of multiple neurons from a single electrode is common practice during extra-cellular recordings. Separation and sorting of spikes originating from the different neurons can be performed either on-line or off-line using multiple methods for pattern matching. However, all spike sorting techniques fail either fully or partially in identifying spikes from multiple neurons when they overlap due to occurrence within a short time interval. This failure, that we termed the ‘shadowing effect’, causes the well-known phenomenon of decreased cross-correlation at zero offset. However, the shadowing effect also causes other artifacts in the auto and cross-correlation of the recorded neurons. These artifacts are significant mainly in brain areas with high firing rate or increased firing synchrony leading to a high probability of spike overlap. Cross correlation of cells recorded from the same electrodes tends to reflect the autocorrelation functions of the two cells, even when there are no functional interactions between the cells. Therefore, the cross-correlation function tends to have a short-term (about the length of the refractory period) peak. A long-term (hundreds of milliseconds to a few seconds) trough in the cross-correlation can be seen in cells with bursting and pausing activities recorded from the same electrode. Even the autocorrelation functions of the recorded neurons feature firing properties of other neurons recorded from the same electrode. Examples of these effects are given from our recordings in the globus pallidus of behaving primates and from the literature. Results of simulations of independent simple model neurons exhibit the same properties as the recorded neurons. The effect is analyzed and can be estimated to enable better evaluation of the underlying firing patterns and the actual synchronization of neighboring neurons recorded by a single electrode.

Introduction

Extracellular recording of neuronal activity is a major tool in neurophysiological studies of the brain. Microelectrodes are used for recording local currents deriving from both spiking activity and local field potentials. A single electrode can potentially pick up signals from multiple cells within a local area (Abeles, 1974, Asanuma, 1989). Many types of studies of the nervous system function demand a separation of the recorded signal into spikes originating from different cells. The spikes recorded from different neurons usually differ in size and shape, thereby enabling their sorting into the different sources (Lewicki, 1998, Harris et al., 2000). Recording and separation of multiple neurons from a single electrode allows examination of the behavior of a population of neurons that are usually much closer than multiple neurons recorded by different electrodes. Such studies are therefore mandatory for understanding local neural networks (Eggermont, 1990, Abeles, 1991).

There are plenty of methods for decomposing the output of single electrodes into parallel spike trains (see review of spike sorting methods in Lewicki, 1998). The methods differ in the algorithms used for sorting (ranging from amplitude discrimination to principal and independent component analysis of the spike shape), the working time frame (real time and offline methods) and the verification methods. In general, all methods suffer from problems common in classical signal detection methods, e.g. false positives (noise in the signal is classified as real spikes) and false negatives (real spikes are rejected as noise). However, when using spike-sorting methods, additional errors may occur. These errors include false match (a spike generated by one unit is classified to a different unit) and double match (a single waveform is classified as belonging to more than one class). The number of such errors can be reduced by better signal to noise recording conditions, and by more careful and elaborate sorting methods.

When the errors in either sorting or identification are systematic, they might cause effects, which seem to derive from the properties of the neurons rather than from the classification procedure (Quirk and Wilson, 1999). Such a systematic classification error occurs due to the spike overlap problem. All sorting methods perform quite well when the spikes recorded from the electrode are sufficiently separated in time. However, when multiple spikes appear closely, causing an overlap of their effects on the recorded signal, all sorting methods perform significantly worse. This overlap may result in several consequences: none of the spikes is identified (complete false negative); only one of the spikes is identified (partial false negative); or the overlapped signal is identified as another different spike (false match). The overlapping problem is usually not handled, although some methods have been developed for reducing the misidentification cause by the overlap (Lewicki, 1998). These methods include identification by neural networks (Chandra and Optican, 1997) and overlap decomposition (Atiya, 1992, Lewicki, 1994, Zouridakis and Tam, 1997). However, whichever methods are used, overlapping spikes are identified significantly worse than well-separated spikes. In this manuscript we show that the auto and cross-correlation functions of simultaneously recorded units (especially in brain areas with high firing rates and synchronized discharge) are significantly affected by the sorting limits. Thus, short and long term synchronization might appear in the cross-correlograms due to the sorting problems in spite of the fact that the neurons fire independently. Finally, we describe methods for estimating these artifacts, and thus enabling better understanding of the firing patterns and synchronization of neighboring neurons in the central nervous system.