Unsupervised automatic online spike sorting using reward-based online clustering
Introduction
One of the basic approaches for understanding brain mechanisms is monitoring the electrical activity of the neurons [1] which helps us to decode the brain information which is vital for BMIs. These neural electrical signals are usually recorded with one or more electrodes, but each electrode records a superposition of activities of several adjacent and farther neurons. The adjacent neurons to the electrode shape a spike train and the farther neurons contribute to the background noise [1]. Hence, spike sorting is required to distinguish each part of the recorded signals and obtain information from the brain [2]. In addition, it has been shown that even a basic spike sorting is beneficial for BMI signal decoding [3].
There are wide variety of spike sorting approaches on neuronal pattern recognition [4]. Most spike sorting algorithms comprise three major steps: spike detection, feature extraction, and clustering [5].
In spike detection, the common method is applying a simple threshold to extract the actual spike waveforms from the raw data [4]. Besides, more complicated techniques, which are based on the wavelet method, such as wavelet transform (WT) [6] or wavelet packets decompositions (WPD) [7] and the energy signal [8,9] for reducing harmful effects on the spike shape and spike detecting, are introduced. In [10] a sparse coding and a compressive sensing method are introduced to solve the problem of overlapping spikes. Moreover, in order to simplify the spike detection step, some studies use mathematical transformation to obtain a new signal from the raw signals that identified spikes in a simple manner [11,12].
Feature extraction is mapping from the original data space to another space, and features are properties of a cluster that specify it from the other clusters. Besides, feature extraction can map the data in a space with lower dimension and as a result, reduce the computational complexity. The other desirable property of feature extraction is that it makes a cluster more distinct, so, it facilitates the clustering process [13]. The most common method for feature extraction and order reduction, which is used in spike sorting literature, is Principal Component Analysis (PCA) [1] which is mostly used in offline clustering methods. Moreover, Wavelet coefficients are proper features to discriminate clusters, but they are complex [6,7]. The other features are the first and second extrema (FSDE) method that are simple shape-based features [13]. There are other features like Diffusion Map features that use a graph of the data for feature extraction [14], and Locality Preserving Projection (LPP) that constructs a graph incorporating neighborhood information of the dataset [15], for offline applications. In contrast, the FSDE features that can be applied for online applications. Recently, a feature extraction approach is introduced that uses shape, phase, and distribution together [16]. It seems that the shape-based feature extractions are more appropriate for online applications rather than other mapping-based methods.
The main part of most spike sorting methods is the clustering technique. Hence, different clustering approaches are presented in the literatures, such as k-means [7,13,14,17], expectation maximization [18,19], superparamagnetic clustering method [6], landmark-based spectral clustering [15], hierarchical adaptive means (HAM) clustering [20], neural networks [8], support vector machines (SVMs) [21]. However, none of the mentioned approaches successfully provided generally optimal results in sorting the spikes. In addition, because of the nature of BMIs, real-time spike sorting is necessary and none of above-mentioned clustering methods satisfy this requirement. Hence, in this study, a new method for online clustering is proposed. This method is Reward-Based Online Clustering (RBOC) which is formed based on the reinforcement learning algorithm. The significant property of this proposed technique is the capability of real-time implementation which is required by BMIs. Besides, this method can automatically detect the clusters while there is no knowledge about the number of clusters.
Reinforcement learning algorithms such as Q-learning have been built based on interaction with the environment and receiving a reward or punishment [22]. Reward or punishment signal can be a physical or real signal, such as a banana given to a monkey for accomplishing a task [23]. It also can be a purely mathematical definition for example distance with the target for an agent. Such algorithms are acting real-time, i.e., when the algorithm starts, information or experience completes by receiving the reward or punishment signal. Because the experience can take place without a supervisor, the algorithm can be done unsupervised. In this case, only the reward or punishment signal definition is required for the algorithm. Besides, this algorithm can be used as an unsupervised automatic real-time clustering method.
Each experience accompanies with a cost and takes plenty of time. Therefore, the concept of planning helps reinforcement learning algorithm to be faster and cheaper. In the planning algorithm, along with the experience, a model of the system is built. The simulated model helps to gain virtual experiences quickly without real interaction with the environment. Then the algorithm would be faster [24].
In this study, the procedure consists of three-parts, detection using simple threshold, feature extraction with Aligned First Derivative (AFD) method, and the new main part which is online unsupervised clustering method with Reward-Based Online Clustering (RBOC) approach.
The evaluation is performed with a set of simulated datasets. Each dataset simulates one trial of recording the brain signal using a single electrode. These sets are classified into four categories based on the degree of complexity regarding the similarity of the spikes in the spike train. Each smaller dataset has four different noise levels. After evaluation, the proposed method is applied to the experimental data obtained from the rat brain using a single electrode.
Section snippets
Simulated data
The simulated datasets first introduced in [7] and were frequently used in previous similar works. The simulated data used in this work consist of 16 different datasets. For Construction of these datasets, a database of 594 different spike waveforms, which are created by averaging of real records from the neocortex and basal ganglia, have been used. Initially, the background noise is generated by several randomly selected spikes from the database with different random amplitudes. This noise is
Results
This section consists of two subsections. First, for a comparative study and validating the proposed method, some simulated (ground truth) datasets are used. Then, the method applied to an experimental dataset obtained from the rat brain.
Discussion
This paper consists of three parts, i.e., spike detection, feature extraction, and clustering. We evaluated the method with a ground truth dataset which is used frequently in the spike sorting literature. The evaluation performed in three sections.
First, just the detection part is examined. The results of this part show that the detection part had only on average about 2.5% false negative and under 2% false positives. It means the detection part accomplished its mission, but there were some
Conclusion
This paper presented a new method for sorting the spikes. It consists of three parts: detection and alignment of spikes, feature extraction, and automatic online clustering.
In the detection and alignment of spikes, a register for storing a set of real-time sampled data was used. Moreover, spikes were detected using two simple thresholds and aligned with their peaks or valleys.
The feature extraction consists of two main parts. First, finding the CPs and calculating the maximum slope on both
Ethical standards
- •
The authors declare that all experiments on human subjects were conducted in accordance with the Declaration of Helsinki http://www.wma.net and that all procedures were carried out with the adequate understanding and written consent of the subjects.
- •
The authors also certify that formal approval to conduct the experiments described has been obtained from the human subjects review board of their institution and could be provided upon request.
- •
If the studies deal with animal experiments, the authors
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (26)
- et al.
Past, present and future of spike sorting techniques
Brain Res. Bull.
(2015) - et al.
A novel automated spike sorting algorithm with adaptable feature extraction
J. Neurosci. Methods
(2012) - et al.
A method for spike sorting and detection based on wavelet packets and Shannon’s mutual information
J. Neurosci. Methods
(2002) - et al.
Feature extraction using first and second derivative extrema (FSDE) for real-time and hardware-efficient spike sorting
J. Neurosci. Methods
(2013) - et al.
Automatic spike sorting by unsupervised clustering with diffusion maps and silhouettes
Neurocomputing.
(2015) - et al.
Spike sorting using locality preserving projection with gap statistics and landmark-based spectral clustering
J. Neurosci. Methods
(2014) - et al.
Hierarchical Adaptive Means (HAM) clustering for hardware-efficient, unsupervised and real-time spike sorting
J. Neurosci. Methods
(2014) - et al.
A Review on Recording of Neural Signal to Monitor Brain Activity
(2019) - et al.
To sort or not to sort: the impact of spike-sorting on neural decoding performance
J. Neural Eng.
(2014) - et al.
Multiple neural spike train data analysis: state-of-the-art and future challenges
Nat. Neurosci.
(2004)
Unsupervised spike detection and sorting with wavelets and superparamagnetic clustering
Neural Comput.
Neural spike sorting under nearly 0-dB signal-to-noise ratio using nonlinear energy operator and artificial neural-network classifier
IEEE Trans. Biomed. Eng.
A new action potential detector using the MTEO and its effects on spike sorting systems at low signal-to-noise ratios
IEEE Trans. Biomed. Eng.
Cited by (10)
Edge computing on TPU for brain implant signal analysis
2023, Neural NetworksAn optimized GMM algorithm and its application in single-trial motor imagination recognition
2022, Biomedical Signal Processing and ControlCitation Excerpt :The task of the clustering algorithm is to find a way that group the objects into sensible clusters by revealing similarities and differences among objects [4]. It was revealed by subdividing the objects into groups, with maximally possible proximity of within-group subset, and minimally possible proximity of between-group subset [5,6]. Clustering algorithms can be divided into hierarchical clustering, partition clustering, density clustering, grid clustering, and model clustering [7].
Nearly symmetric orthogonal wavelets for time-frequency-shape joint analysis: Introducing the discrete shapelet transform's third generation (DST-III) for nonlinear signal analysis
2021, Communications in Nonlinear Science and Numerical SimulationCitation Excerpt :Thus, spike sorting (SS) [18,25,26] based on minimum-distortion strategies has emerged as one of the most relevant data analysis problems in nonlinear science research. Neurocomputation [27], biomedicine [28], signal control [29], among others, represent the main fields where SS is of paramount importance. Over the years, different automated strategies [30] have been adopted to efficiently implement SS algorithms, such as phase- and shape-based [31], filtering-based [32], direct feature extraction- [33,34] and fusion-based [35], neural network-based [36], Gaussian Mixture Model-based [37], just to mention a few possibilities, where the specific objectives vary considerably [38–40].
Wireless closed-loop deep brain stimulation using microelectrode array probes
2024, Journal of Zhejiang University: Science BFrom End to End: Gaining, Sorting, and Employing High-Density Neural Single Unit Recordings
2022, Frontiers in Neuroinformatics