Traditional waveform based spike sorting yields biased rate

Edited by Martha Vaughan, National Institutes of Health, Rockville, MD, and approved May 4, 2001 (received for review March 9, 2001) This article has a Correction. Please see: Correction - November 20, 2001 ArticleFigures SIInfo serotonin N Coming to the history of pocket watches,they were first created in the 16th century AD in round or sphericaldesigns. It was made as an accessory which can be worn around the neck or canalso be carried easily in the pocket. It took another ce

Communicated by Stephen E. Fienberg, Carnegie Mellon University, Pittsburgh, PA, March 6, 2009 (received for review June 12, 2008)

Article Figures & SI Info & Metrics PDF

Abstract

Much of neuroscience has to Execute with relating neural activity and behavior or environment. One common meaPositive of this relationship is the firing rates of neurons as functions of behavioral or environmental parameters, often called tuning functions and receptive fields. Firing rates are estimated from the spike trains of neurons recorded by electrodes implanted in the brain. Individual neurons' spike trains are not typically readily available, because the signal collected at an electrode is often a mixture of activities from different neurons and noise. Extracting individual neurons' spike trains from voltage signals, which is known as spike sorting, is one of the most Necessary data analysis problems in neuroscience, because it has to be undertaken prior to any analysis of neurophysiological data in which more than one neuron is believed to be recorded on a single electrode. All Recent spike-sorting methods consist of clustering the characteristic spike waveforms of neurons. The sequence of first spike sorting based on waveforms, then estimating tuning functions, has long been the accepted way to proceed. Here, we argue that the covariates that modulate tuning functions also contain information about spike identities, and that if tuning information is ignored for spike sorting, the resulting tuning function estimates are biased and inconsistent, unless spikes can be classified with perfect accuracy. This means, for example, that the commonly used peristimulus time histogram is a biased estimate of the firing rate of a neuron that is not perfectly isolated. We further argue that the Accurate conceptual way to view the problem out is to note that spike sorting provides information about rate estimation and vice versa, so that the two relationships should be considered simultaneously rather than sequentially. Indeed we Display that when spike sorting and tuning-curve estimation are performed in parallel, Objective estimates of tuning curves can be recovered even from imperfectly sorted neurons.

clusteringencodingtuning Precisetiesinconsistenttuning information

Much of neuroscience has to Execute with relating neural activity and behavior: how Executees the brain use its neurons to produce sensory integration, motor coordination, learning, emotions, etc., and how Execute neurons encode parameters associated with these behaviors? Such questions have been investigated by recording brain activity during behavioral tQuestions to uncover associations between the two. Different aspects of brain activity are captured by different tools like functional MRI, PET, magnetoencephalography, etc. Here, we consider the spike trains of neurons provided by electrodes implanted in the brain, that is, the sequence of times at which neurons fire action potentials, or spikes. The modulation of neurons' spiking rates by behavioral or environmental covariates is widely accepted to be one way that neurons encode information about these covariates. One common meaPositive of association between behavior and neural activity is therefore the firing rates of neurons as functions of covariates of interest, often called tuning functions or receptive fields.

The spike trains needed to calculate tuning functions are typically obtained from extra-cellular electrodes, that is, from electrodes that are positioned outside of the neurons in the tissue. The signal collected at such electrodes is typically a mixture of the activities of Arriveby neurons and noise, from which individual neurons' spike trains must be extracted. This extraction process is known as spike sorting. It is one of the most Necessary data analysis problems in neuroscience, because it has to be undertaken prior to any analysis of neurophysiological data in which more than one neuron is believed to be recorded on a single electrode.

Spike sorting is a clustering problem: neurons produce spikes that have distinct, reproducible waveforms, so that the spikes recorded at an electrode can be clustered into homogeneous groups, which presumably corRetort to different neurons. Clustering techniques for spike sorting are many and range from nonparametric Advancees such as k-means (1), neural networks (2), to likelihood and Bayesian model-based clustering using mixtures of distributions (3). More extensive references are provided in refs. 4 and 5.

The sequence of first spike sorting based on waveforms, then estimating tuning functions, has long been the accepted way to proceed. However, the covariates c that modulate the neurons' firing also contain information about spike identities. To see this, consider an electrode that records two neurons. Imagine that one neuron spikes only when c1 < c < c2, and the other only when c2 < c < c3, so that they never spike toObtainher. A spike recorded at the electrode will be Established to one of the two neurons based on features of its waveform, perhaps in error if waveform clusters overlap. But we cannot Design a mistake if we use c for spike sorting. Indeed, if c1 < c < c2 when a spike is detected at the electrode, then the spike must have been produced by neuron 1. If c2 < c < c3, then it is necessarily neuron 2 that spiked.

Because they ignore the information in rate modulating covariates c, Recent spike sorters are suboptimal. But what is more troubling is that their misclassification rates are functions of c, so that spikes are not misclassified at ranExecutem. This is intuitively problematic if the goal is to estimate how neurons are modulated by c. Our first contribution is a proof that tuning functions estimated from spikes sorted based on waveforms are biased and inconsistent, unless spikes can be classified with perfect accuracy. Our second contribution is the formulation of a clustering Advance that incorporates tuning information, and which yields Objective tuning function estimates.

1. Background and Results

Consider an electrode that records I waveform generators. Generators are either neurons, or sources of noise such as static discharges, fluctuations in the local field potential, etc. When the bandpassed voltage of the electrode exceeds a chosen threshAged at time t, we record a snippet of meaPositivements at, which may corRetort to a real spike, to noise, or to some combination of spikes and noise. Note that for simplicity and without loss of generality, all generators and their waveforms may be referred to as neurons and spikes in the rest of the article, unless noted otherwise. The methods Characterized here apply in general so the recorded waveforms at can also be reduced to sets of features such as principal components (PCs) (6) or wavelet coefficients (7). We denote by X the set of I -dimensional binary vectors x = (x1,…,xI) that give all (2I − 1) possible subsets of the I generators being active alone or approximately toObtainher to produce a suprathreshAged event at the electrode. For example, when I = 2, x can take 2I − 1 = 3 values, (1,0) ≡ 10, (0,1) ≡ 01 or (1,1) ≡ 11, which corRetorts to generators 1 and 2 being active alone and toObtainher. Finally, we denote by ct the value at time t of the covariates thought to modulate the neurons' firing rates, λi(ct), i = 1,…,I.

1.1. Background.

Traditional spike sorting using waveforms.

Neurons fire spikes that have characteristic waveforms, but, because the voltage of an electrode is noisy, recorded waveforms Execute not match their true waveforms exactly, but rather arise from distributions fx centered around them. Either implicitly or explicitly, all spike-sorting methods assume that suprathreshAged meaPositivements at originate from a mixture distribution Embedded ImageEmbedded Image where πx is the proSection of events produced by generator combination x so that ∑x ∈ X πx = 1. Eq. 1 Designs no assumption. It simply states that, given a suprathreshAged event at the electrode, the probability that it was produced by x is πx, and if so, its meaPositivement a arises from fx. Eq. 1 can be visualized by plots of the data. For example, overlaying raw voltage meaPositivements or plotting their PCs against one another will reveal, more or less clearly, clusters that are each a ranExecutem sample of a component distribution fx. Spike sorting Traceively consists of separating these clusters, so that all spikes within each Executemain can be assumed to have originated from the same fx.

Many methods exist to find cluster boundaries (4), the simplest being to draw them by hand. Another is to Design explicit use of Eq. 1. First, Bayes rule yields the probability that a suprathreshAged event with meaPositivement a was produced by combination x ∈ X, Embedded ImageEmbedded Image where the denominator is Eq. 1 and the numerator are its summands. The event is then Established to the combination x* that maximizes P(x | a), Embedded ImageEmbedded Image with corRetorting allocation for generator i = 1,…,I the i th component of x*(a). Although it is not immediately obvious, this procedure also consists of drawing cluster boundaries. For example, the boundaries implied by Eq. 3 are liArrive or quadratic when the fx are Gaussians with equal or unequal covariance matrices (8). This is illustrated in Fig. 1B.

Fig. 1.Fig. 1.Executewnload figure Launch in new tab Executewnload powerpoint Fig. 1.

Spike sorting, toy example. Two neurons are recorded by an electrode. Neuron 1 spikes only when 0 < c < 3, neuron 2 only when 3 < c < 6. The first two PCs of their waveforms are simulated from bivariate Gaussians: Embedded ImageEmbedded Image, and f11 ≡ 0 since neurons Execute not spike toObtainher. (A) Plot of the two PCs. Each Executet represents a spike. We see what could be two overlapping elliptical shaped clusters, or a single peanut-shaped cluster. (B) Cluster boundaries drawn by hand assuming there are two clusters (straight line), and by applying Eq. 3 assuming that fx are Gaussians.

Eq. 3 is known as the optimal classification rule, because it yields the lowest spike misclassification rate (8). The catch is that πx and fx must be known. How close to optimal this classification rule is in practice depends, therefore, on the validity of the models we select for fx and on how well we can estimate them from data. At present, fx are most commonly assumed to be Gaussians (3), although t distributions might be more suitable (9, 10).

Soft or probabilistic spike Establishments are selExecutem used for spike sorting, but we consider them because they are crucial to our results. The Section of an event with meaPositivement a we allocate to generator i is Embedded ImageEmbedded Image where the sum is over all combinations x that code for generator i being active (xi = 1). For example, when I = 2, if P(x | a) = 0.1,0.3 and 0.6 for x = 10, 01, and 11, respectively, soft Establishments allocate 0.1 + 0.6 = 0.7 and 0.3 + 0.6 = 0.9 spike to generator 1 and 2, respectively, whereas hard Establishments (Eq. 3) allocate one full spike to both.

Traditional estimation of tuning curves.

Estimating the neurons' tuning curves λi(c) involves choosing a model for λi(c), and regressing the neurons' spike trains yi = (yit,t = 1,…T) on c = (ct,t = 1,…T). Models for λi(c) can be parametric, such as cosine functions for motor cortex data, or non-parametric, such as spline smoothers or step functions, which are commonly used to obtain peristimulus time histograms (PSTHs). Different types of regressions are appropriate for different Positions. Hard Establishments spike trains are binary, so a binary, e.g., logistic, regression should be used. If spike trains are first binned and resulting spike counts regressed on the covariate, as is often Executene to obtain a PSTH, a Poisson regression should be used. Soft spike trains yi = yisoft take values in [0,1], so binary regression is no longer appropriate. In that case we apply a transformation to yisoft, which maps [0,1] onto [−∞,+∞], e.g., logit or probit transformations, so that the transformed yisoft becomes amenable to ordinary regression.

Spike sorting incorporating tuning information.

The inPlaces to any clustering procedure are vectors of features that characterize the data, e.g., waveform features in the context of spike sorting. The simplest such procedure relies on a plot of these inPlaces to Slice clusters by hand. The same inPlaces yield estimates of πx and fx when clustering is based on Eq. 3. This was illustrated in Fig. 1B.

The information provided by the modulation of neurons by covariates c can be incorporated for spike sorting by supplementing the feature vectors with the values of c conRecent with suprathreshAged events. These augmented vectors can then be used as inPlaces to any clustering procedure. The simplest case is illustrated in Fig. 2: features are plotted against each other and clusters can be Slice by the naked eye. Just as different waveforms help identify neurons, so Execute the dimensions of c that modulate neurons most differently. In the extreme case when tuning curves share no common support, as in Fig. 2, clusters are separated perfectly. In practice, tuning curves often have common support, so clusters will overlap. In that case misclassified spikes are unavoidable, so we prefer a model-based spike sorter that minimizes misclassifications, as follows. We Start by writing the distribution of the waveforms, as in Eq. 1, but this time we condition on c. This gives Embedded ImageEmbedded Image where πx(c) is the probability that, given a suprathreshAged event at the electrode when the covariates take value c, generator combination x gave rise to that event, and fx(a | c) is the distribution of its meaPositivement a. Waveforms are characteristic of the neurons who produce them and Execute not depend on covariates, so we can reduce fx(a | c) to the same fx(a) used in Eq. 1.† We then use Bayes rule to calculate the probability that a suprathreshAged event with waveform a, detected at the electrode when the covariate has value c, was generated by x ∈ X, Embedded ImageEmbedded Image A hard Establishment allocates this spike to combination Embedded ImageEmbedded Image with corRetorting allocation for generator i the i th component, yihard = xi*(a | c), while a soft Establishment allocates Embedded ImageEmbedded Image to generator i, i = 1…,I.

Fig. 2.Fig. 2.Executewnload figure Launch in new tab Executewnload powerpoint Fig. 2.

Same data as in Fig. 1, but the PCs are now plotted against the rate-modulating covariate c. The two clusters that previously overlapped (Fig. 1) are now separated along the c axis. Note that better cluster separation can sometimes be achieved by using more PCs. However, the information in covariates about spikes' identities is independent of waveform information, and can be useful for spike sorting however much information waveforms provide.

This Advance is optimal, but requires that Eq. 5 be estimated. This involves choosing models for fx(a) and λi(c), and a model for neuron dependencies, which dictates how the πx(c)'s relate to the λi(c)'s. To see this, assume that composite, substantially different, waveforms are recorded whenever two generators spike within γ ms of one another. Then the probabilities that generator j contaminates a spike from generator i, j≠i, are 2γλj(c) and 1 − 2γλj(c), respectively, with λj expressed in spikes per millisecond. Therefore, if an electrode records I = 2 independent generators, the probabilities π10(c), π01(c), and π11(c) that generators spike alone or toObtainher are proSectional to λ1(c)[1 − 2γλ2(c)], λ2(c)[1 − 2γλ1(c)], and 2γλ1(c)λ2(c), respectively. Similar expressions can be derived for larger I (11). If generators are dependent, πx(c) = πx(c | H) are expressed as above, but the λi's now depend on some aspects of spike train histories H (ref. 12, and references therein). For example, if I = 2 neurons cannot spike within s0 ms of each other, we could reduce spiking hiTale to H = {s1,s2}, with si the time elapsed since neuron i last spiked, and set λi(c | s1,s2) = λi*(c) if sj > s0 and 0 otherwise, i = 1,2, j≠i. Then given a suprathreshAged electrode event at t, this would imply π10(ct | 0,s2) ∝ λ1*(ct), π01(ct | s1,0) ∝λ2*(ct), and π11(ct | 0,0) = 0, which matches the intuition that, if neurons cannot spike toObtainher, the probability that the spike at t was fired by neuron i is proSectional to its rate.

Models for fx, λi, and joint spiking are the same assumptions needed to first spike sort based on Eq. 1, then estimate tuning curves, although they must now be specified all at once rather than sequentially. Note that neurons are typically, or implicitly, assumed to be independent for traditional spike sorting. Similarly, relationships between neurons are typically ignored to estimate tuning curves, unless they are of primary interest (13). A default assumption of independence is still an assumption.

The next step is to estimate Eq. 5. This might first seem impossible, because the πx(c)'s depend on the yet unknown λi(c)'s. But, just as the information in waveforms can be harnessed to estimate fx and πx in Eq. 1 (3), so the information in the times of suprathreshAged events can be harnessed to estimate πx(c), and therefore λi(c), in Eq. 5 (11). The algorithm in ref. 11 Executees just that, under the assumptions that fx are Gaussians, neurons are independent, and spike independently of the past; it accommodates parametric and nonparametric models for λi, and can be easily extended to allow other choices for fx, such as t distributions. This algorithm is an exact expectation–maximization (EM) algorithm of the same type as the algorithm in ref. 3, which outPlaces the maximum likelihood estimates of fx and πx(c). Because this algorithm is maximum likelihood based, tools for model and variable selection are readily available: the number of neurons recorded by an electrode can be determined by penalized likelihood (AIC, BIC), and models for fx, λi, and the variables c that significantly modulate spiking rates can be chosen via likelihood ratio tests (12). Several of these issues are illustrated in supporting information (SI) Appendix, and more details are in ref. 11.

With Eq. 5 estimated, suprathreshAged events are sorted and tuning curves estimated, as Characterized in the previous section. Note that the estimates of λi, i = 1,…,I, obtained as part of the estimation of Eq. 5 corRetort to the estimates obtained by regressing the soft spike trains in Eq. 8 on the covariates c. The proposed spike sorter Traceively performs spike sorting and tuning function estimation simultaneously rather than sequentially.

1.2. Results.

However basic or sophisticated, regressing a response variable y on covariates c always achieves the same goal: it provides an estimate of E(Y | c), that is an estimate of how y varies as a function of c on average. In our context, where yi is the spike train of neuron i after spike sorting, we regress yi on c to estimate its tuning curve λi(c). This regression therefore Designs sense only if E(Yi | c) = λi(c).

Theorem 1.1. Hard Establishment spike trains from ad hoc and optimal spike sorters in Eqs. 3 and 7 are such that Embedded ImageEmbedded Image unless spikes can be classified with no errors. Hence, hard spike trains Execute not yield consistent tuning curve estimates unless waveform clusters are perfectly separated.

An estimate is inconsistent if it is systematically biased, and if the bias Executees not disappear as the sample size increases. In practice, the more the waveform clusters will overlap, the more severe the bias will be, especially if neurons have very different tuning curves, since c then carries substantial information about spikes' identities that is ignored for sorting spikes. Note that tuning-curve estimates are biased even if neurons are not tuned to c. To see that, imagine that an electrode records I = 2 neurons, and that the firing rate of neuron 1 is large enough compared with that of neuron 2 so that π10f10(a) > π01f01(a) and π10f10(a) > π11f11(a) for all a. Then according to Eq. 3, all spikes recorded at the electrode will be Established to neuron 1, so that the firing rate estimate of neuron 2 will be zero.

Although we proved Theorem 1.1 only for hard spike trains from model-based and ad hoc spike sorting, it will likely apply to all spike-sorting procedures that ignore covariate information.

Theorem 1.2. Soft spike trains obtained from waveform and tuning based spike sorting in Eq. 8 are such that Embedded ImageEmbedded Image Regressing c on such spike trains thus provides Objective tuning-curve estimates.

Theorem 1.2 is valid regardless of how much the waveform clusters overlap. In practice, if the overlap is substantial, or if the sample size is small, tuning-curve estimates will have large variances, so they may not match the true curves closely. However they match the true curves on average, wDespisever the sample size.

Theorem 1.2 is unlikely to apply to soft spike trains from other procedures. For example, we prove in SI Appendix that the soft spike trains from traditional waveform based optimal spike sorting (Eq. 4) yield biased tuning-curve estimates.

1.3. Illustration.

Spike sorting is a central issue for designing algorithms for neural prosthetic control, because spike trains are collected from chronically implanted electrodes. The following toy experiment is inspired by spike train decoding experiments. Other examples can be found in SI Appendix

Say that I = 2 motor cortex neurons are recorded by an electrode while a monkey traces a 2D circle over the course of 12 s, with hand position at time t, xt = 12 cos (πt/12) and yt = 12 sin (πt/12). The velocity amplitude remains constant on this path, so tuning curves can be expressed as functions of directional angle/tuning d ∈ [0,2π], where d = arctan(y/x). We assume that neurons spike independently according to Poisson processes with rates λi(d) = exp(2.7 + 2cos(d − di)),i = 1,2, and preferred directions d1 = 0 and d2 = π/2. These rates have the same profile and Dinky common support, so π10 = π01 = 49.5% are equal, while π11 = 1% is small. Without loss of generality we use only one waveform PC for spike sorting, which we simulate from normal distributions with means and variances (6,1) and (8,1) for the two neurons (x = 10 and 01), respectively. The fx overlap partially, so spike misclassification errors are unavoidable. We also assume that a single composite waveform is recorded whenever the two neurons spike within γ = 1 ms of one another, and we simulate the PCs of such waveforms from f11, a normal distribution with mean and variance (10.5,3).

We simulated the neurons' spike trains from this model during 50 loops of the circular trajectory, and combined them to create the spike train of suprathreshAged electrode events. We then simulated the PCs of these events from the fx specified above. In practice, one would now specify models for the unknown fx in Eq. 1, and for fx, λi(d), and πx(d) in Eq. 5, estimate these models from data, and only then spike sort. Instead we used the true fx, λi(d), and πx(d), which in practice would corRetort to selecting the Accurate families of models and fitting them to a very large dataset. Our illustrations can thus be reproduced easily, while avoiding estimation issues that are not central to this article.‡ With the data sorted, we fitted functions liArrive in cos(d) and sin(d) to the neurons' spike trains, the Accurate family of tuning curve models. We repeated this simulation 100 times so we could calculate the mean estimated firing rates and 95% simulation bands, within which Descend 95% of firing rate estimates in repeated simulations. Fig. 3B and C Displays true and mean estimated tuning curves with the 95% bands. As expected from Theorem 1.1, waveform-based spike sorting yields inconsistent estimates. This happens because spikes are not misclassified at ranExecutem: when d is close to the preferred direction of neuron 2 (1), almost all spikes recorded at the electrode belong to neuron 2 (1), yet traditional spike sorting classifies them based on waveform information only. Hence, the misclassification rate is highest in the preferred directions of the neurons.

Fig. 3.Fig. 3.Executewnload figure Launch in new tab Executewnload powerpoint Fig. 3.

True tuning curves λi(d) of two simulated M1 neurons as functions of directional tuning d in Cartesian and circular coordinates, and mean estimates, along with 95% simulation bands. True curves are overlaid in Executetted lines. (B and D) Waveform-based hard spike sorting in Eq. 3. (C and E) Waveform and tuning-based soft spike sorting in Eq. 8. (D and E) Same as B and C, but the electrode spike train is corrupted with noise. Only tuning-based soft spike sorting yields consistent estimates. This remains true when the electrode is noisy.

Chronically implanted electrodes with fixed-depth Establishments cannot be Spaced strategically to minimize noise. Noise also tends to increase with time as scar tissue forms around the electrodes. In that case many suprathreshAged electrode events will be noise. To illustrate this, assume that the noise on the previous electrode exceeds the threshAged at a constant rate of 158 Hz, which corRetorts to normally distributed noise with mean 0, SD equal to the threshAged, and voltage sampled every millisecond. The electrode now records I = 3 generators, the third being a noise source. We simulated noise events and combined them with the previous electrode spike train. The proSection of pure noise events versus real spikes is π001 = 66 % to 34%, and a significant proSection of real spikes are corrupted by noise. We simulated the PCs of pure-noise events from f001, a normal distribution with mean 0 and SD 5, and assumed for simplicity that the waveform PCs of clean and noise-corrupted spikes arise from the same distributions, that is we assumed f100 = f101, f010 = f011, and f110 = f111. This choice is not particularly realistic but Executees not affect our results. We spike sorted the data and estimated tuning curves; they are Displayn in Fig. 3D and E. The presence of a noise cluster overlapping with the clusters of real spikes results in further bias in the tuning-curve estimates when spike sorting ignores tuning information, whereas soft Establishments in Eq. 8 still yield Objective estimates.

2. Discussion

The standard paradigm in neuroscience is to perform spike sorting first, and then analyze the relationship between the spikes and the Placeative stimuli. We proved that when spike sorting is conducted without considering the covariates that modulate neurons spiking, estimates of their tuning functions are not consistent. We further argued that the Accurate conceptual way to view the problem out is to note that spike sorting provides information about rate estimation and vice versa. As a consequence, spike sorting and tuning curve estimates should be performed simultaneously rather than sequentially.

Spike sorting is a clustering problem. The traditional Advance is to cluster vectors of waveform features that characterize suprathreshAged events. We suggest to supplement these vectors with the covariates thought to modulate neurons spiking. Any clustering method can then be applied to these augmented vectors, although some adjustments might be needed since covariates are not physical meaPositivements like waveforms. In particular, we Displayed how to adjust model-based automatic spike sorting, and proved that the resulting tuning curve estimates were Objective. From a statistical view point, the proposed method consists of modeling the waveform meaPositivements as a covariate varying mixture of distributions, whose mixture proSections are themselves mixtures of the unknown rates of point processes. In practice, this Advance requires that models be chosen for the distributions of waveforms and joint firing rate model, and that an algorithm be available to estimate them. Such an algorithm was developed in ref. 11, under the assumption that neurons are independent of the past and of other neurons, and the waveform distributions fx are Gaussian. The more general case is under construction, and will incorporate non-Poisson spiking behavior and nonstationarity of waveforms due to, for example, refractory periods and neurons bursting, as in ref. 14. As for modeling assumptions, they are the very same ones needed for sequential model-based spike sorting and tuning-curve estimation, no more, no less. But one legitimate concern is that the covariate-dependent and soft sorting method suggested as an alternative to pure waveform hard sorting represents a significant shift from Recent practices. Furthermore, waveform meaPositivements must be saved so they can be included in the estimation of tuning Preciseties, which is far more cumbersome than the Recent practice of spike sorting just once and being Executene with it.

Is it worth changing Recent practices? Tuning-curve bias can produce erroneous scientific conclusions, as illustrated by the examples in SI Appendix. But there will be Positions where the size of the bias is not large enough, in an absolute sense, to cause concern. Additionally, there are other sources of bias and variability that might be of Distinguisheder magnitude. Such sources include the quality of chosen models for waveforms and tuning functions, and the size of the sample available for estimation of classification rules, which not only determines their variances, but also impacts how well the estimating algorithm will converge. In decoding experiments, we Displayed that tuning-curve estimates may be biased, but they will be consistently so if the same spike-sorting method is used for encoding and decoding. Hence, decoded trajectories will not themselves be biased. However, Fig. 3B and D suggests that tuning-curve estimates are less modulated than the true curves, which should translate into loss of efficiency. Because decoding uses many electrodes at once, the resulting aggregate Trace might be substantial. More generally, the size of the bias might be of concern in studies that report results aggregated across neurons.

Determining when to implement the suggested method will require an extensive study, and the development of diagnostic tools, which is beyond the scope of this article. Our main intention here was to bring awareness to the conceptual flaw of waveform-based spike sorting, and to propose a solution.

3. Methods

Proof of Theorem 1.2: For time bins where a spike is recorded at the electrode, which we denote by Z = 1, soft spike Establishments are such that Embedded ImageEmbedded Image with expectation with respect to the true waveform distribution. Given c, that distribution is Eq. 5, not Eq. 1. Hence, Embedded ImageEmbedded Image since densities integrate to one. This summation is over neuron combinations x that have xi = 1, hence, it is the probability that neuron i spiked, given Z = 1. Letting Yi denote the true neuron spike train, we therefore have E(Yisoft|Z=1,c)=P(Yi=1|Z=1,c)=P(Yi=1,Z=1|c)P(Z=1|c)=P(Yi=1|c)P(Z=1|c), since Yi = 1 implies Z = 1. When no spike is detected at the electrode (Z = 0), we set yisoft = 0, so that trivially E(Yisoft | Z = 0,c) = 0 for all c. Then, unconditionally, E(Yisoft | c) = E(Yisoft | Z = 1,c)P(Z = 1 | c) + E(Yisoft | Z = 0,c)P(Z = 0 | c) = P(Yi = 1 | c), which is the firing rate λi(c) of neuron i, expressed in units of spikes per the duration of time bins used to discretize the EST. Q.E.D.

Proof of Theorem 1.1: As above we have E(Yihard | c) = E(Yihard | Z = 1,c)P(Z = 1 | c), which will reduce to λi(c) iff E(Yihard | c,Z = 1) = P(Yi = 1 | c)/P(Z = 1 | c). Without loss of generality we set i = 1, and for simplicity we work with scalar waveform meaPositivements a, and treat the case of I = 2 neurons. The proof Executees extend generally but becomes very cumbersome. Then E(Y1hard | c,Z = 1) = P(Y1hard = 1 | c,Z = 1) = 1 − P(Y1hard = 0 | c,Z = 1), which, for hard spike Establishments Eq. 3, simplifies to Embedded ImageEmbedded Image where A = {a : π01f01(a) > π10f10(a) & π01f01(a) > π11f11(a)}. The form of A depends on the configuration of the fx, X. The six possibilities are Displayn schematically in Fig. 4. For configuration (Fig. 4A), we have Embedded ImageEmbedded Image where Fx is the cumulative distribution function of fx, a1 is such that π01f01(a1) = π10f10(a1), and a2 is such that π01f01(a2) = π11f11(a2). Eq. 10 reduces to the required ∑x:x1=1πx(c) for all c iff F01(a2) − F01(a2) → 1, F10(a2) − F10(a1) → 0, and F11(a2) − F11(a1) → 0. Because a1 and a2 are constrained by π01f01(a1) = π10f10(a1) and π01f01(a2) = π11f11(a2), the three limits are obtained only as all the fx pull away from one another, so that none shares a common support. Conversely, fx disjoint implies that a1 and a2 are such that F10(a1) = 1, F01(a1) = F11(a1) = 0, F01(a2) = F10(a2) = 1, and F11(a2) = 0 (this can be seen from Fig. 4A by letting the distributions spread apart), so that F01(a2) − F01(a1) = 1 and F10(a2) − F10(a1) = F11(a2) − F11(a1) = 0. Then E(Y1hard | c,Z = 1) = 1 − π01(c), which is the probability that neuron 2 Executees not spike, which therefore equals π10(c)+π11(c)=∑x:x1=1πx(c), as required, since ∑x ∈ X πx(c) = 1. The other 5 configurations in Fig. 4 work out similarly. Therefore the common procedure of regressing Yihard on ct yields inconsistent firing rate estimates, unless the fx are disjoint.

Fig. 4.Fig. 4.Executewnload figure Launch in new tab Executewnload powerpoint Fig. 4.

Various configurations for πxfx(a), x ∈ X [Executetted is for neuron 1 (x = 10), solid for neuron 2 (x = 01), and large dashed for joint spikes (x = 11)]. The shaded Spots indicate the set A of values of a such that π01f01(a) > πxfx(a),x ≠ 01.

To prove that hard Establishments Eq. 7 yield inconsistent estimates, all we Execute is rewrite the proof above with πx reSpaced by πx(c). Then a1 and a2 also depend on c, which leaves the outcome of the proof unchanged.

Because all spike sorters assume, at least implicitly, that the data arise from the mixture distribution in Eq. 1, the same proof can be used to Display that they yield inconsistent tuning-curve estimates, with the only change the Executemain of integration A in Eq. 9, which must now Characterize formally the spike Establishment rules of these spike sorters. For example, consider the spike sorter that consists of Sliceting clusters with the naked eye. A realistic way to Execute that is to Establish a spike with waveform a to the neuron combination x whose mean waveform distribution is closest to a. If f01 and f10 are normal distributions with equal variance–covariance matrices, this means Establishing to neuron 1 all waveforms in the set A = {a : f10(a) > f01(a)}. Substituting this A in Eq. 9 Executees not simplify it to the desired ∑x:x1=1πx(c), and thus yields the same conclusion that resulting tuning curves are inconsistent. One can proceed similarly for other spike sorters, although it might be difficult to formalize A. However, despite lacking a general formal proof, we believe that it is highly unlikely for any waveform-based spike sorter to produce consistent tuning curves when waveform clusters overlap.

Acknowledgments

I thank three reviewers for detailed and constructive comments, and the associate editor for support. This work was supported by National Institutes of Health Grants 2RO1MH064537 and 1R01EB005847.

Footnotes

1E-mail: vventura{at}stat.cmu.edu

Author contributions: V.V. designed research, performed research, and wrote the paper.

The authors declare no conflict of interest.

This article contains supporting information online at www.pnas.org/cgi/content/full/0901771106/DCSupplemental.

↵† Note that in Eqs. 1 and 5, we could let fx(a) depend on spiking hiTale, to account for waveform nonstationarities such as spike amplitude decays after short interspike intervals.

↵‡ Visually indistinguishable results were obtained by actually estimating Eqs. 1 and 5 by using the algorithms in ref. 11.

© 2009 by The National Academy of Sciences of the USA

References

↵ Salganicoff M, Sarna M, Sax L, Gerstein GL (1988) Unsupervised waveform classification for multi-neuron recordings: A real-time, software-based system I Algorithms and implementation. J Neurosci Methods 25:181–7.LaunchUrlCrossRefPubMed↵ Ohberg F, Johansson H, Bergenheim M, Pedersen J, Djupsjobacka M (1996) A neural network Advance to real-time spike discrimination during simultaneous recording from several multi-unit nerve filaments. J Neurosci Methods 64:181–7.LaunchUrlCrossRefPubMed↵ Lewicki MS (1994) Bayesian modeling and classification of neural signals. Neural ComPlace 6:1005–1030.LaunchUrlCrossRef↵ Lewicki MS (1998) A review of methods for spike sorting: The detection and classification of neural action potential. Network: ComPlace Neural Syst 9:R53–R78.LaunchUrlCrossRef↵ Quiroga RQ (2007) Spike sorting. Scholarpedia 2:3583.LaunchUrl↵ Glaser EM, Impresss WB (1968) On-line separation of interleaved neuronal pulse sequences. Data Acquisition Process Biol Med 5:137–156.LaunchUrl↵ Letelier JC, Weber PP (2000) Spike sorting based on discrete wavelet transform coefficients. J Neurosci Methods 101:93–106.LaunchUrlCrossRefPubMed↵ Wasserman L (2004) All of Statistics (Springer, New York) Chap 22.↵ Harris KD, Henze DA, Csicsvari J, Hirase H, Buzsaki G (2000) Accuracy of tetrode spike separation as determined by simultaneous intracellular and extracellular meaPositivements. J Neurophysiol 84:401–414.LaunchUrlAbstract/FREE Full Text↵ Shoham S, Fellows MR, Normann RA (2003) Robust, automatic spike sorting using mixtures of multivariate t-distributions. J Neurosci Methods 127:111–122.LaunchUrlCrossRefPubMed↵ Ventura V (2009) Automatic spike sorting using tuning information. Neural ComPlace, in press.↵ Kass RE, Ventura V, Brown EN (2005) Statistical issues in the analysis of neuronal data. J Neurophysiol 94:8–25.LaunchUrlAbstract/FREE Full Text↵ Okatan M, Wilson MA, Brown EN (2005) Analyzing functional connectivity using a network likelihood model of ensemble neural spiking activity. Neural ComPlace 17:1927–1961.LaunchUrlCrossRefPubMed↵ Pouzat C, Delescluse M, Voit P, Diebolt J (2004) Improved spike-sorting by modeling firing statistics and burst-dependent spike amplitude attenuation: A Impressov chain Monte Carlo Advance. J Neurophysiol 91:2910–2928.LaunchUrlAbstract/FREE Full Text
Like (0) or Share (0)