An unsupervised real-time spike sorting system based on optimized OSort

Unsupervised automatic online spike sorting using reward-based online clustering

Transmitting large amounts of data sensed by multi-electrode array devices is widely considered to be a challenging problem in designing implantable neural recording systems. Spike sorting is an important step to reducing the data bandwidth before wireless data transmission. The feasibility of spike sorting algorithms in scaled CMOS technologies, which typically operate on low frequency neural spikes, is determined largely by its power consumption, a dominant portion of which is leakage power. Our goal is to explore energy saving architectures for online spike sorting without sacrificing performance. In the face of non-stationary neural data, training is not always a feasible option. We present a low-power architecture for real-time online spike sorting that does not require any training period and has the capability to quickly respond to the changing spike shapes.

Brain Machine Interface (BMI) systems demand real-time spike sorting to instantaneously decode the spike trains of simultaneously recorded cortical neurons. Real-time spike sorting, however, requires extensive computational power that is not feasible to implement in implantable BMI architectures, thereby requiring transmission of high-bandwidth raw neural data to an external computer. In this work, we describe a miniaturized, low power, programmable hardware module capable of performing this task within the resource constraints of an implantable chip. The module computes a sparse representation of the spike waveforms followed by "smart" thresholding. This cascade restricts the sparse representation to a subset of projections that preserve the discriminative features of neuron-specific spike waveforms. In addition, it further reduces telemetry bandwidth making it feasible to wirelessly transmit only the important biological information to the outside world, thereby improving the efficiency, practicality and viability of BMI systems in clinical applications.

Feature extraction is a critical step in real-time spike sorting after a spike is detected. Features should be informative and noise insensitive for high classification accuracy. This paper describes a new feature extraction method that utilizes a feature denoising filter to improve noise immunity while preserving spike information. Six features were extracted from filtered spikes, including a newly developed feature, and a separability index was applied to select optimal features. Using a set of the three highest-performing features, which includes the new feature, this method can achieve spike classification error as low as 5% for the worst case noise level of 0.2. The computational complexity is only 11% of principle component analysis method and it only costs nine registers per channel.

Much development has been directed toward improving the performance and automation of spike sorting. This continuous development, while essential, has contributed to an over-saturation of new, incompatible tools that hinders rigorous benchmarking and complicates reproducible analysis. To address these limitations, we developed SpikeInterface, a Python framework designed to unify preexisting spike sorting technologies into a single codebase and to facilitate straightforward comparison and adoption of different approaches. With a few lines of code, researchers can reproducibly run, compare, and benchmark most modern spike sorting algorithms; pre-process, post-process, and visualize extracellular datasets; validate, curate, and export sorting outputs; and more. In this paper, we provide an overview of SpikeInterface and, with applications to real and simulated datasets, demonstrate how it can be utilized to reduce the burden of manual curation and to more comprehensively benchmark automated spike sorters.

Developing electrophysiological recordings of brain neuronal activity and their analysis provide a basis for exploring the structure of brain function and nervous system investigation. The recorded signals are typically a combination of spikes and noise. High amounts of background noise and possibility of electric signaling recording from several neurons adjacent to the recording site have led scientists to develop neuronal signal processing tools such as spike sorting to facilitate brain data analysis. Spike sorting plays a pivotal role in understanding the electrophysiological activity of neuronal networks. This process prepares recorded data for interpretations of neurons interactions and understanding the overall structure of brain functions. Spike sorting consists of three steps: spike detection, feature extraction, and spike clustering. There are several methods to implement each of spike sorting steps. This paper provides a systematic comparison of various spike sorting sub-techniques applied to real extracellularly recorded data from a rat brain basolateral amygdala. An efficient sorted data resulted from careful choice of spike sorting sub-methods leads to better interpretation of the brain structures connectivity under different conditions, which is a very sensitive concept in diagnosis and treatment of neurological disorders. Here, spike detection is performed by appropriate choice of threshold level via three different approaches. Feature extraction is done through PCA and Kernel PCA methods, which Kernel PCA outperforms. We have applied four different algorithms for spike clustering including K-means, Fuzzy C-means, Bayesian and Fuzzy maximum likelihood estimation. As one requirement of most clustering algorithms, optimal number of clusters is achieved through validity indices for each method. Finally, the sorting results are evaluated using inter-spike interval histograms.

BackgroundPrincipal component analysis (PCA) has been widely employed for automatic neuronal spike sorting. Calculating principal components (PCs) is computationally expensive, and requires complex numerical operations and large memory resources. Substantial hardware resources are therefore needed for hardware implementations of PCA. General Hebbian algorithm (GHA) has been proposed for calculating PCs of neuronal spikes in our previous work, which eliminates the needs of computationally expensive covariance analysis and eigenvalue decomposition in conventional PCA algorithms. However, large memory resources are still inherently required for storing a large volume of aligned spikes for training PCs. The large size memory will consume large hardware resources and contribute significant power dissipation, which make GHA difficult to be implemented in portable or implantable multi-channel recording micro-systems.MethodIn this paper, we present a new algorithm for PCA-based spike sorting based on GHA, namely stream-based Hebbian eigenfilter, which eliminates the inherent memory requirements of GHA while keeping the accuracy of spike sorting by utilizing the pseudo-stationarity of neuronal spikes. Because of the reduction of large hardware storage requirements, the proposed algorithm can lead to ultra-low hardware resources and power consumption of hardware implementations, which is critical for the future multi-channel micro-systems. Both clinical and synthetic neural recording data sets were employed for evaluating the accuracy of the stream-based Hebbian eigenfilter. The performance of spike sorting using stream-based eigenfilter and the computational complexity of the eigenfilter were rigorously evaluated and compared with conventional PCA algorithms. Field programmable logic arrays (FPGAs) were employed to implement the proposed algorithm, evaluate the hardware implementations and demonstrate the reduction in both power consumption and hardware memories achieved by the streaming computingResults and discussionResults demonstrate that the stream-based eigenfilter can achieve the same accuracy and is 10 times more computationally efficient when compared with conventional PCA algorithms. Hardware evaluations show that 90.3% logic resources, 95.1% power consumption and 86.8% computing latency can be reduced by the stream-based eigenfilter when compared with PCA hardware. By utilizing the streaming method, 92% memory resources and 67% power consumption can be saved when compared with the direct implementation of GHA.ConclusionStream-based Hebbian eigenfilter presents a novel approach to enable real-time spike sorting with reduced computational complexity and hardware costs. This new design can be further utilized for multi-channel neuro-physiological experiments or chronic implants.

A large amount of neurophysiological research is based on the study of biological neural populations. The data is gathered from extra-cellular recordings with multi-electrode arrays (MEAs). The signal is a stream containing an unknown number of neural sources superpositioned with non-stationary noise components. In order to analyze the recorded spike trains, a prior separation into its individual components is required. The increasing number of sensors on a MEA surface demand an automatic spike sorting procedure. In this article the proposed spike sorting method replaces the manual steps with artificial neural networks to enable an unsupervised separation of mixed signal components. Furthermore the artificial neural network based feature extraction allows realtime processing and ensures high flexbibility and adaptivity for unknown signals with non-stationary noise components.

During the analysis of neurobiological signals acquired during neurosurgical procedures such as Deep Brain Stimulation, one focuses mainly on the signal’s background and spiking activity. Spikes, the electrical impulses generated by neuron cells, are events lasting about 1.5 ms with voltage typically well below 200 \(\upmu \)V. As the shape of spike derives from the physical construction of the neuron’s cell membrane and its distance from the recording point, one can infer that spikes of different shapes were generated by different neuron cells. This information can be beneficial during the analysis of the acquired signals, and the procedure of grouping spikes according to their shapes, is called spike sorting. Typically in this procedure, there are defined various attributes characterizing spike shape, and subsequently, the standard clustering method is used to produce groups of spikes of different shapes. This may cause some difficulties that arise both from the definition of good discriminating attributes as well as from the fact that one does not know beforehand how many clusters of shapes to expect. The approach presented in this paper is different, using the fully unsupervised and explainable AI, it generates groups of spike shapes in a fully automatic way. No attributes have to be defined nor the expected number of shape clusters provided. Spike waveforms are fed into the autoencoder that has defined additional constraints on its latent layers. After training, data obtained from those layers can be used for the purpose of spike shape clustering. Beside shape class, one also receives information on how much given spike shape differs from the mean shape of its class.

This paper proposes a novel k-means algorithm for clustering analysis for the cases that the true number of clusters in a data or points set is not known in advance. That is, assuming that the number of seed-points in the algorithm is set to be larger than the true number k of clusters in the data set, the proposed algorithm can assign the k seed-points to the actual clusters, respectively, with the extra seed-points corresponding to the empty clusters, i.e., having no winning points according to a newly defined distance. Via using the Mahalanobis distance, the proposed algorithm can be further extended to elliptical clustering analysis. It is demonstrated well by the experiments on simulated data set and the wine data that the proposed k- means algorithm can find the correct number of clusters in the sample data with a good correct classification rate. Moreover, the algorithm is successfully applied to unsupervised color image segmentation.

The marked degree of heterogeneity in persons with developmental dyslexia has motivated the investigation of possible subtypes. Attempts have proceeded both from theoretical models of reading and the application of unsupervised learning (clustering) methods. Previous cluster analyses of data obtained from persons with reading disabilities have suffered from the inherent limitations of unsupervised learning methods. Specifically, the reliability and stability of cluster solutions have proven difficult to determine. Recent developments in the clustering literature have addressed these concerns by permitting checks on the internal validity of the solution. Resampling methods produce consistent groupings of the data independent of initialization effects, while the gap statistic provides a confidence measure for the determination of the optimal number of clusters present in the data. Combining these methods produces a robust data-driven classification that can be compared with theoretically based subtypes to produce better-informed models of developmental dyslexia. The present study is a novel application of resampling (bootstrap aggregating or bagging) methods and the gap statistic to the subtyping of children with developmental dyslexia. The specific aims of this study are: (1) to illustrate the use of bagging methods and the gap statistic in multivariate data obtained from children with developmental dyslexia; and (2) to compare the bagged clustering thresholded by the gap statistic against the predictions of the double-deficit hypothesis. The double-deficit hypothesis is a prominent theoretical model of developmental dyslexia, which predicts three subtypes: phonological, rate, and phonological-rate impaired readers. Three simulated data sets with known cluster structure were created to check the validity and illustrate the utility of the bagged clustering with the gap statistic in data with known structure. Subsequently, a clinical database of standardized test data (eight tests) from 93 children with developmental dyslexia was clustered using these methods. This procedure was repeated on a database of 93 children without reading disability matched for gender and age as a control. Finally, the clustering was repeated on the entire database of 186 participants. Cluster solutions were obtained for an increasing number of clusters (1-10) and were tested against the null hypothesis that no subtypes were present, i.e. the data represented a single cluster. Four clusters were identified in the children with developmental dyslexia. There was no evidence of significant cluster structure in the children without dyslexia. Two clusters were identified when children with and without reading impairments were considered together. Among the participants with developmental dyslexia, there was evidence of a phonological-deficit cluster, a rapid-naming cluster, and a cluster showing both depressed phonological processing and rapid naming. These accounted for 73 of the 93 participants (78%). All three are predicted by the double-deficit hypothesis. The fourth cluster consisted of children with normal phonological and rapid naming ability incommensurate with their high verbal ability. An analysis of variance with post-hoc multiple comparisons demonstrated that the phonological, rapid-naming, and double deficit clusters did not differ significantly in age, but the fourth cluster was comprised of significantly older children. The mixed data set revealed two clusters. One cluster consisted almost entirely of the double-deficit and phonological subtypes. The other consisted of the participants without dyslexia and the children with dyslexia demonstrating either a single rapid naming deficit or standardized test scores in the normal range. A silhouette analysis indicated that the four-cluster solution for the children with developmental dyslexia was superior to the two-cluster solution obtained for the entire data set. The study provides support for the presence of distinct subtypes in children with developmental dyslexia and for the double-deficit hypothesis. Specifically, this study finds three subtypes predicted by the double-deficit hypothesis without the assumption of an a priori theoretical model of reading. Taken together, these subtypes account for 79% of the participants with dyslexia. Further, the percentages of children in each subtype are in good agreement with previous studies. The participants in the subtype not predicted by the double-deficit hypothesis were significantly older than the other three groups. Recent advances in unsupervised learning can be expected to aid the improvement and refinement of the definition of developmental dyslexia. If reliable and consistent subtypes can be identified among persons with developmental dyslexia, it is reasonable to assume that diagnostic and intervention efforts will be greatly improved.

On-implant spike sorting methods employ static feature extraction/selection techniques to minimize the hardware cost. Here we propose a novel framework for real-time spike sorting based on dynamic selection of features. We select salient features that maximize the geometric-mean of between-class distances as well as the associated homogeneity index effectively to best discriminate spikes for classification. Wave-shape classification is performed based on a multi-label window discrimination approach. An external module calculates the salient features and discrimination windows through optimizing a replica of the on-implant operation, and then configures the on-implant spike sorter for real-time online operation. Hardware implementation of the on-implant online spike sorter for 512 channels of concurrent extra-cellular neural signals is reported, with an average classification accuracy of ~88%. Compared with other similar methods, our method shows reduction in classification error by a factor of ~2, and also reduction in the required memory space by a factor of ~5.

The emergence of multi-electrode array enables the study of real-time neurophysiological activities across multiple regions of the brain. However, the real-time extracellular action potentials recorded on any electrode represent the simultaneous electrical activity of an unknown number of neurons which present a critical challenge to the accuracy of interpretation and identification of the neural circuitry in the subsequent analysis. In this paper, we present a principal component analysis approach utilizing Hebbian eigenfilter to identify the corresponding electrical activities of each neuron, namely spike sorting. The Hebbian eigenfilter greatly simplifies the computational complexity of eigen-projection. An efficient FPGA-based Hebbian eigenfilter is proposed. The performance, accuracy and power consumption of our Hebbian eigenfilter are thoroughly evaluated through synthetic spike trains. The proposal enables real-time spike sorting and analysis, and leads the way towards future motor and cognitive neuroprosthetics.

WeDIV – An improved k-means clustering algorithm with a weighted distance and a novel internal validation index

The end user often needs to define extremely complex interpretation tasks and to require the analysis results to be quantitatively proven for each pixel without the test data. To this end, this paper extends the ideas underlying the model-based unsupervised classification method previously proposed by us (Koltunov and Ben-Dor 2001). Consistently with that method, the quality of assigning a pixel to a cluster is defined as the lower confidence bound (l.c.b.) of the corresponding posterior probability estimate. We propose to compute the l.c.b.s in an approximate way using the Fisher information matrix instead of the bootstrap scheme suggested previously, leading to an l.c.b.-estimation procedure that is faster by a factor of hundreds to thousands, while being reasonably accurate. The issue of selecting the number of clusters is considered in accordance with the quantitative requirements for the level of detail and the reliability of the thematic interpretation. Specifically, the l.c.b.s form a novel selection criterion that allows the most detailed landscape descriptions to be provided with at least a pre-specified value of confidence. This implies detection of highly overlapping clusters, leading to very detailed segmentations. Consequently, numerous thematic classification and object detection problems can be solved, based on single clustering and assuming that thematic classes are unions of components. Then the thematic classification accuracy can be computed in a well-founded manner for each separate pixel, using the obtained covariance matrices of the posterior probability estimators of component membership. The procedures for thematic mapping and object detection are described. The accuracy of the l.c.b. estimation and stability of the new criterion in choosing the number of clusters are illustrated on simulated datasets. The hyperspectral data analysis experiment performed demonstrates part of the developments described in this paper. Several issues that are relevant for remote sensing data interpretation are addressed constructively. In particular, we draft the novel algorithms that use model-based cluster analysis for detection and recognition of remotely sensed objects based on prior information on their size and shape. In addition, we introduce a generalized approach to unsupervised feature extraction from data acquired by a plurality of sensors of different physical nature. A new data model generalizing the traditional Gaussian mixture model is also presented.