Toward Automated Electrode Selection in the Electronic Depth Control Strategy for Multi-unit Recordings

K.W. Wong et al. (Eds.): ICONIP 2010, Part II, LNCS 6444, pp. 17–25, 2010. © Springer-Verlag Berlin Heidelberg 2010

Toward Automated Electrode Selection in the Electronic Depth Control Strategy for Multi-unit Recordings

Gert Van Dijck1, Ahmad Jezzini2, Stanislav Herwik3, Sebastian Kisban3, Karsten Seidl3, Oliver Paul3, Patrick Ruther3, Francesca Ugolotti Serventi2, Leonardo Fogassi2,4, Marc M. Van Hulle1, and Maria Alessandra Umiltà2,4

1 Computational Neuroscience Research Group, Laboratorium voor Neuro- en Psychofysiologie, Katholieke Universiteit Leuven, Herestraat 49, 3000 Leuven, Belgium 2 Department of Neuroscience, University of Parma, Via Volturno 39, 43100 Parma, Italy 3 Microsystem Materials Laboratory, Department of Microsystems Engineering (IMTEK),

University of Freiburg, Georges-Koehler-Allee 103, 79110 Freiburg, Germany 4 Italian Institute of Technology, Section of Parma, Parma, Italy

[email protected], [email protected], [email protected], [email protected]

Abstract. Multi-electrode arrays contain an increasing number of electrodes. The manual selection of good quality signals among hundreds of electrodes becomes impracticable for experimental neuroscientists. This increases the need for an automated selection of electrodes containing good quality signals. To motivate the automated selection, three experimenters were asked to assign quality scores, taking one of four possible values, to recordings containing action potentials obtained from the monkey primary somatosensory cortex and the superior parietal lobule. Krippendorff’s alpha-reliability was then used to verify whether the scores, given by different experimenters, were in agreement. A Gaussian process classifier was used to automate the prediction of the signal quality using the scores of the different experimenters. Prediction accuracies of the Gaussian process classifier are about 80% when the quality scores of different experimenters are combined, through a median vote, to train the Gaussian process classifier. It was found that predictions based also on firing rate features are in closer agreement with the experimenters’ assignments than those based on the signal-to-noise ratio alone.

Keywords: Continuous wavelet transform, Electronic depth control, Gaussian process classifier, Inter-rater reliability, Multi-unit recordings, Spike detection.

1 Introduction

Multi-electrode arrays (MEAs) are able to monitor the simultaneous spiking activity of many neurons [1] and are, therefore, in a position to provide valuable insights into how multiple neurons process information and how different brain areas interact. MEAs typically consist of tens of electrodes with inter-electrode distances ranging from 100 to 500 μm [2]. Recent advances in CMOS-based microprobes [3], support

18 G. Van Dijck et al.

in-vivo recordings up to 500 electrodes with inter-electrode distances as small as 40.7 μm [3], hence, at a spatial range of an individual neural cell’s soma diameter. Signal quality can degrade over time due to apoptosis, tissue drift, relaxation, inflammation and reactive gliosis, among other reasons. Hence, there is a need for (re)selecting or (re)positioning the electrodes. Essentially 2 different systems can be used for obtaining and maintaining a good signal quality among different electrodes: movable microelectrodes [4-7] and the electronic depth control system [3]. The movable microelectrode systems mechanically position single electrodes independently, whereas the electronic depth control system switches electronically between different microelectrodes on a single shaft (up to 500 electrodes [3]) once the array is implanted. In the electronic depth control system, 8 electrodes, out of the total number of electrodes, can be read simultaneously from a single shaft. This is due to technological limitations in the switching electronics and the lithography. So far, the electrode selection was performed manually by the experimenter or semi-automatically. In the latter case, electrodes are sorted according to the signal-to-noise ratio (SNR) [3]. To fully automate the electrode selection, it is of interest to know whether experimenters can agree on the quality of a signal. This is studied in section 3 using Krippendorff’s alpha inter-rater reliability. So far, the SNR was used as the quality metric in both the movable microelectrode systems [4,6-7] and the electronic depth control [3]. It is studied in section 4 whether signal features other than the SNR may be needed to reflect the experimenter’s assignments of a quality score to signals containing action potentials. This is performed by studying the accuracy of a Gaussian process classifier (GPC) [8] as a function of the SNR feature and firing rate features.

2 Experiments

The neural activity used in this study was collected from two series of recording sessions performed using a new generation microprobes [9], semi-chronically implanted in the cortex of an awake macaque monkey.

The silicon-based probes applied in this study comprise four slender, 8-mm-long probe shafts arranged as a comb, as shown in Figure 1(a). Each probe shaft has a width and thickness of 140 µm and 100 µm, respectively, and carries nine circular Pt electrodes with a diameter of 35 µm (cf. Figure 1(a)). Out of these 36 electrodes, the first eight electrodes of each shaft, i.e. 32 electrodes, as counted from the electrode tip are accessible for the recording experiments. The inter-electrode pitch and the distance between the probe shafts were set to 250 µm and 550 µm, respectively. The two-dimensional electrode array as shown in Figure 1(a) is fabricated using microsystem technology detailed elsewhere [9]. Figure 1(b) shows the silicon-based probe assembled to a U-shaped polyimide cable comprising the interconnection part for a zero insertion force (ZIF) connector. The probe insertion into the brain tissue is performed using the insertion device in Figure 1(b). During insertion, the probe and its cable are fixed on the insertion device using vacuum suction. They are released for operating of the probe after retraction of the insertion device. The two semi-chronic recording sessions were performed using the same neural device. In the first series of recording sessions, the neural device was implanted in the primary somatosensory cortex (SI) and it was kept in the cortex for 8 days.

Toward Automated Electrode Selection in the Electronic Depth Control Strategy 19

Fig. 1. (a) Silicon-based probe comb with four slender shafts with planar electrodes and a connector pads for cable assembly and (b) probe comb assembled to a U-shaped ribbon cable temporarily mounted onto an insertion device using vacuum for insertion

The second implantation was located in the superior parietal lobule (SPL) and the neural device was kept in the cortex for 10 days. The acquisition time of each trial lasted 4 seconds that corresponded to 2 seconds before and 2 seconds after the somatosensory stimulation. In order to assess the quality of the recorded signal from the implanted neural device, the neuronal activity of each electrode was carefully investigated on-line by the experimenter. In addition to the on-line recorded signal assessment, three experimenters performed an off-line quality signal assessment.

0 1 2 3 4

-0.5

0

0.5

time [s]

Am

plitu

de

Score 4

0 1 2 3 4

-0.5

0

0.5

time [s]

Am

plitu

de

Score 3

0 1 2 3 4-0.4

-0.2

0

0.2

time [s]

Am

lpitu

de

Score 2

0 1 2 3 4-0.4

-0.2

0

0.2

time [s]

Am

plitu

de

Score 1

Fig. 2. Example signals of the four different scores. All 3 raters agreed independently on the score of each of the 4 signals. The stimulus was presented starting from 2 seconds.

This assessment was done separately and independently by each experimenter in order to avoid any possible bias in the evaluation of the neural signal quality. The assessment was done by a visual inspection of the neuronal discharge in single trials, on the basis of the personal experience of the experimenter. The assignment was based on a four levels scale, from 1 to 4. Each number corresponded to a different quality of the recorded neuronal activity, see Figure 2. Score 1 was assigned when no

20 G. Van Dijck et al.

activity was recorded, score 2 when there was only background activity, score 3 when many different spikes were recorded simultaneously but the isolation of a single neuron was not possible. Score 4 was assigned when well isolated spikes were recorded.

3 Inter-rater Reliability

The degree to which raters agree on a quality score for signals, can be computed by an inter-rater reliability coefficient. Hereto, we used Krippendorff’s agreement coefficient alpha (α) [10]. This reliability coefficient can be interpreted as: the degree to which independent observers, here using quality scores on a scale of 1 to 4, respond identically to each individual signal. Properties of α that are important in our case are the ability to deal with more than 2 raters, the use of confidence intervals, the correction for small sample sizes and the possibility to penalize more heavily larger disagreements between raters. To illustrate the latter: a 1-4 confusion between raters, i.e. one rater gives the lowest score 1 and another gives the highest score 4, is not as forgivable as 1-2 confusion or a 2-3 confusion between raters.

The alpha reliability can be computed as:

o

e

D1

Dα = − , (1)

where Do is the observed disagreement and De is the expected disagreement. Do and De are computed as:

2o ck ck

c k

1D o

n= δ∑∑ , (2)

2e c k ck

c k

1D n n

n(n 1)= ⋅ δ

− ∑∑ . (3)

Here, n is the total number of values that have been assigned over all raters, i.e. n is equal to #raters x #signals (here 3 raters x 256 signals). Further, ock is the observed number of c-k pairs, i.e. the number of times one rater gave quality ‘c’ and another

quality ‘k’. 2ckδ is the penalty factor for a disagreement that one rater rated a signal as

quality ‘c’ and another as quality ‘k’ ( 2ckδ = 0 for c = k), nc and nk are the number of

times score ‘c’ and score ‘k’ have been assigned. For details on the computation of the observed coincidences ock we refer to chapter 11 of [10].

The α value can range between -1 and 1, with a -1 indicating systematic disagreement between raters and 1 indicating systematic agreement. A small value of α around 0 indicates that raters do not agree more or disagree less than expected by chance level. In Table 1, we report the reliability scores for different choices of the

penalty function 2ckδ .

Toward Automated Electrode Selection in the Electronic Depth Control Strategy 21

Table 1. Krippendorff’s alpha reliability scores for different metric values. 1000 bootstraps were taken to obtain the 95% confidence interval. α123 is the reliability coefficient computed between all 3 raters.

Metric 2ckδ α123 coefficient 95% confidence interval

Nominal: 2ck

2ck

1,c k

0,c k

δ = ≠

δ = = 0.6458 [0.5989,0.6908]

Ordinal: see [10] 0.8882 [0.8691,0.9048]

Interval: ( )22ck c kδ = − 0.9006 [0.8863,0.9144]

Ratio: 2

2ck

c k

c k

−⎛ ⎞δ = ⎜ ⎟+⎝ ⎠ 0.8658 [0.8435,0.8859]

The first column of Table 1 shows that the nominal metric punishes every

mismatch between the scores of raters in an equally severe way, while in the other metrics the penalty depends on the degree of mismatch. The ordinal metric is likely the most suitable metric here, because it considers the values of the raters as a rank. It is observed that the nominal metric leads to the smallest reliability value, while the other metrics are about the same. This suggests that raters disagree sometimes on the exact quality score, but their assignments will seldom differ by more than 1 value. Larger confusions almost never happen. Indeed, in only 6 out of the 256 signals, a difference in quality larger than 1 was observed between any 2 observers. Under the ordinal (0.8882), interval (0.9006) and ratio (0.8658) metric, the assignments can be considered as reliable.

4 Prediction of Quality Scores

Our goal is now to predict the quality scores assigned by the different raters, since these predictions can then be used in the electronic depth control. Hereto, we train and test a Gaussian process classifier using features extracted from the neural signals. The features are based on the spikes which are detected with a continuous wavelet transform. We consider the following features: (1) the signal-to-noise ratio (SNR) which has been considered so far as the standard in electrode positioning [4,6-7], and selection [3], (2) the maximal firing rate in 20 ms bins around stimulus presentation and (3) the average firing rate in the same interval around stimulus presentation.

First, spikes are detected off-line in the neural signals using a continuous wavelet transform [11]. We used the Daubechies 2 wavelet, detecting spikes that are about 0.5 ms wide. We tried 2 detection thresholds: one equal to 0 and a more conservative one that only detects the larger spikes, see Figure 3. A higher detection threshold will in general lead to a lower probability of detection (PD) and to a lower probability of false alarms (PFA). For an interpretation of the detection thresholds the reader is referred to [11].

22 G. Van Dijck et al.

2348 2350 2352 2354 2356 2358 2360 2362 2364

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

time [ms]

Am

plitu

de

spike detection

Fig. 3. The circles locate the spikes detected with a continuous wavelet transform (Daubechies 2 mother wavelet) using a threshold equal to 0, the diamonds are spike locations with a more conservative threshold equal to 0.2. The boxes are centered on the detected spikes (spikes detected at the diamonds) taking 0.5 ms before and 0.5 ms after the spike location. The average root-mean-square (RMS) of the signal within the boxes (spikes) is divided by the standard deviation outside the boxes (which are the noise segments) in the computation of the signal-to-noise ratio.

The SNR is computed as the average root-mean-square (RMS) of the spikes, N in total, divided by the median absolute deviation (MAD) noiseσ̂ of the noise segments

(see also Figure 3):

N

nn 1

dB 10noise

1RMS(spike (t))

NSNR 20 log

ˆ== ⋅

σ

∑.

(4)

The median absolute deviation was used as a robust estimator for the noise [11]. The SNR quantifies how large spikes are compared to the background noise. However, it has to be noted that only a few, but large spikes can result in a high SNR. Experimenters often are not only interested in obtaining large spikes, but also in how responsive the cell is towards a stimulus. Therefore, also firing rate features were computed. For the analyses performed in the present study, we considered 1 second before and one second after the somatosensory stimulation, in order to be sure to include the whole neuronal response, since the discharge onset varied in relation to the different properties of the neurons recorded from different cortical layers and cortical areas. We segmented this interval into bins of 20 ms, the firing rate of the bin with the highest firing rate was considered as a feature. Furthermore, the average firing rate over the same interval was computed as the third feature.

We used a Gaussian process classifier (GPC) [8] to predict the quality scores. The variational Bayesian approach [8], using a radial basis function (RBF) kernel, was used. The kernel hyperparameters as well as the other parameters of the GPC are found by maximizing the variational lower bound on the marginal likelihood [8]. The advantage of using this GPC compared to some support vector machines is that no extra cross-validation cycles are required for tuning the kernel hyperparameters, they are inferred elegantly in a Bayesian way from the data [8]. We used the leave-one-out

Toward Automated Electrode Selection in the Electronic Depth Control Strategy 23

method for validation. Before training and testing the GPC, the SNR and firing rate features were standard normalized. The classification test accuracies for each rater separately, are shown in Figure 4. The SNR and the bin with maximal firing rate were used to train and test the GPC for the results shown in Figure 4.

rater 1 rater 2 rater 3 raters combined0

20

40

60

80

100

accu

racy

[%]

Primary Somatosensory Cortex and Superior Parietal Lobule

threshold 0.0threshold 0.2

Fig. 4. Classification accuracies for the prediction of the quality scores using the variational Bayesian GPC and a leave-one-out validation. The last pair of bars, labeled ‘raters combined’, was obtained by taking the median vote of all 3 raters for each signal and training and testing the GPC on this median vote. For this median vote the accuracy is about 80%, which is higher than the accuracies of each rater separately.

To study the effect of different feature combinations, we tested (1) the SNR only, (2) the SNR and Max. Fr (bin with maximal firing rate), (3) the SNR and Avg. Fr (average firing rate) and (4) the SNR, Max. Fr and Avg. Fr. The results are shown in Figure 5.

SNR SNR - Max Fr SNR - Avg Fr SNR - Max Fr - Avg Fr0

20

40

60

80

100

accu

racy

[%]

Primary Somatosensory Cortex and Superior Parietal Lobule

threshold 0.0threshold 0.2

Fig. 5. Accuracies of different parameter combinations in the variational Bayesian GPC. The median vote of all 3 raters was used to provide the quality scores to the GPC. Clearly, the accuracies are higher when the SNR is combined with the bin containing the maximal firing rate (the SNR-Max Fr pair of bars) or with the average firing rate (the SNR-Avg Fr pair of bars) compared to the SNR alone (the SNR pair of bars).

24 G. Van Dijck et al.

Comparing previous tables, we observe that the predictions of the quality scores

become more accurate when a firing rate parameter is used in combination with the SNR (Table 3, Table 4 and Table 5), compared to the case when the SNR is used alone (Table 2). The most accurate result is obtained when the SNR is combined with the average firing rate (Table 4): 82.03%. Comparing Table 2 and Table 4, we can make following observations for the most important class (score 4). The recall for class 4 in Table 2 (SNR) is equal to 51/56 = 0.911 which is the same as for the other tables. However, in Table 2 the precision for class 4 is lower, 51/66 = 0.773, compared to 51/54 = 0.944 for Table 4. Hence, the predictions for class 4 become much more precise if besides the SNR also the average firing rate parameter is used. Similar conclusions can be drawn when comparing Table 3 or Table 5 with Table 2: in both cases the predictions become more precise when the SNR is combined with firing rate features.

5 Conclusion

Quality score assignments by experimental neuroscientists seldom differ more than 1 quality value, on a scale of 4 possible values. This was reflected in the high (inter-rater) Krippendorff’s alpha reliability of about 0.88 using the ordinal metric. In previous research, the signal-to-noise ratio (SNR) was used as the quality metric in

Table 2. Confusion matrix SNR

Manually assigned scores

74.22

% score

1 score

2 score

3 score

4

score 1 107 17 1 0

score 2 2 4 4 1

score 3 2 20 28 4

Pre

dict

ed

scor

es

score 4 7 6 2 51

Table 4. Confusion matrix SNR – Avg Fr

Manually assigned scores

82.03

% score

1 score

2 score

3 score

4

score 1 112 19 1 2

score 2 6 25 9 1

score 3 0 3 22 2

Pre

dict

ed

scor

es

score 4 0 0 3 51

Table 3. Confusion matrix SNR – Max Fr

Manually assigned scores

80.08

% score

1 score

2 score

3 score

4

score 1 111 19 1 0

score 2 6 24 11 2

score 3 0 4 19 3 Pre

dict

ed

scor

es

score 4 1 0 4 51

Table 5. Conf. matrix SNR – Max Fr – Avg Fr

Manually assigned scores

79.69

% score

1 score

2 score

3 score

4

score 1 110 19 1 0

score 2 6 24 11 2

score 3 0 4 19 3 Pre

dict

ed

scor

es

score 4 2 0 4 51

Toward Automated Electrode Selection in the Electronic Depth Control Strategy 25

movable microelectrode systems [4,6-7]. In this research a variational Bayesian Gaussian process classifier was used to predict the quality scores, leading to an accuracy of 82% if the SNR and the average firing rate are used together. These results suggest that the selection of the electrodes that capture signals of the highest quality can be performed by using the predictions of the quality scores.

Acknowledgments. We thank Richard Csercsa, Peter Pazmany Catholic University, Budapest, for taking part in assigning quality scores to the neural signals. This research was performed within the framework of the Information Society Technologies (IST) Integrated Project NeuroProbes of the 6th Framework Program (FP6) of the European Commission (Project number IST-027017). Gert Van Dijck and Marc M. Van Hulle are sponsored by the CREA financing program (CREA/07/027) of the K.U. Leuven and the Belgian Fund for Scientific Research – Flanders (G.0588.09).

References

1. Brown, E.N., Kass, R.E., Mitra, P.P.: Multiple Neural Spike Train Data Analysis: State-of-the-Art and Future Challenges. Nat. Neurosci. 7, 456–461 (2004)

2. Pine, J.: A History of MEA Development. In: Taketani, M., Baudry, M. (eds.) Advances in Network Electrophysiology: Using Multi-Electrode Arrays, pp. 3–23. Springer-Verlag New York Inc., New York (2006)

3. Seidl, K., Torfs, T., De Mazière, P.A., Van Dijck, G., Csercsa, R., Dombovari, B., et al.: Control and Data Acquisition Software for High-density CMOS-based Microprobe Arrays Implementing Electronic Depth Control. Biomed. Tech. 55, 183–191 (2010)

4. Sato, T., Suzukia, T., Mabuchi, K.: A New Multi-electrode Array Design for Chronic Neural Recording, with Independent and Automatic Hydraulic Positioning. J. Neurosci. Methods 160, 45–51 (2007)

5. Fee, M.S., Leonardo, A.: Miniature Motorized Microdrive and Commutator System for Chronic Neural Recording in Small Animals. J. Neurosci. Methods 112, 83–94 (2001)

6. Cham, J.G., Branchaud, E.A., Nenadic, Z., Greger, B., Andersen, R.A., Burdick, J.W.: Semi-chronic Motorized Microdrive and Control Algorithm for Autonomously Isolating and Maintaining Optimal Extracellular Action Potentials. J. Neurophysiol. 93, 570–579 (2005)

7. Jackson, N., Sridharan, A., Anand, S., Baker, M., Okandan, M., Muthuswamy, J.: Long-term Neural Recordings using MEMS based Movable Microelectrodes in the Brain. Front. Neuroeng. 3(10), 1–10 (2010)

8. Girolami, M., Rogers, S.: Variational Bayesian Multinomial Probit Regression with Gaussian Process Priors. Neural Comput. 18, 1790–1817 (2006)

9. Herwik, S., Kisban, S., Aarts, A., Seidl, K., Girardeau, G., Benchenane, K., et al.: Fabrication Technology for Silicon Based Microprobe Arrays used in Acute and Subchronic Neural Recording. J. Micromech. Microeng. 19, 074008 (2009)

10. Krippendorff, K.: Content Analysis: an Introduction to its Methodology. Sage Publications Inc., California (2004)

11. Nenadic, Z., Burdick, J.W.: Spike Detection Using the Continuous Wavelet Transform. IEEE Trans. Biomed. Eng. 52, 74–87 (2005)