User Manual Privacy Policy Disclaimer Contact us
  Advanced SearchBrowse





Real-Time Analysis of Extracellular Multielectrode Recordings

There are no MPG-Authors available
Fulltext (public)
There are no public fulltexts stored in PuRe
Supplementary Material (public)
There is no public supplementary material available

Franke, F. (2011). Real-Time Analysis of Extracellular Multielectrode Recordings. PhD Thesis, Technische Universität Berlin, Berlin, Germany.

Cite as: http://hdl.handle.net/11858/00-001M-0000-0013-B8E0-0
For the understanding of how neural networks grow, learn and are able to fulfill their impressive functions, a reliable way to monitor their activity is crucial. The neuroscientist has a steadily growing toolbox for this purpose with one tool being of particular importance even though it is also one of the oldest techniques applied, namely extracellular recordings. The main reasons for its broad usage are probably the low cost and relative ease of application: an electrode needs to be placed in the vicinity of a neuron. Since the most important way of neurons to communicate is by means of electrical events called action potentials, extracellular recordings provide the possibility to "listen" to the communication in a neural network. A major drawback of that technique, however, is the complexity of the analysis of the recordings that is caused by low signal quality as well as the measurement inherent simultaneous recording of not only one but many neurons whose signals need to be separated. Despite the long lasting history of extracellular recordings a satisfactory solution to its associated problem of detection and correct classification of single neuronal action potentials - called spike sorting - is still not readily available. Especially with the massive increase of its usage and the development of more sophisticated electrode arrays that are in principle able to record over long time periods from hundreds of neurons simultaneously, extracellular recordings will also play a major role for future neuroscience, brain machine interfaces and medical applications. But these applications make great demands on the algorithms used: because of the huge amount of extracellular recordings humans will not be able to supervise the spike sorting process any longer so the algorithms need to be fully automatic. For so called closed-loop experiments, where the spike sorting result will be e.g. used to stimulate other neurons in real-time the spike sorting procedures also have to provide the results in real-time. The quickly increasing number of recording electrodes will force the methods to optimally combine information from all available electrodes and deal with a higher computational burden. And finally, in long lasting experiments slow changes in the recording setup like changes in electrode position will cause non-adaptive algorithms to fail. This work is the investigation of the applicability of linear filters for the purpose of fast automatic and adaptive spike detection and sorting. Linear filters provide the advantage of easy implementation - also in hardware - are computationally fast and embedded in a well developed theory. The calculations in this work are done in the discrete-time signal space which simplifies the derivations and shows interesting connections of the spike sorting problem to optimal matched filters, beamforming, space-time adaptive processing, time series forecasting and Wiener filtering. The role of the noise covariance matrix that arises naturally in these domains and is also crucial for many - especially linear-filter based - spike sorting algorithms is investigated. Based on those filters two spike sorting algorithms are proposed that are suited for real-time implementation, can adapt to non-stationary data and make optimal use of multielectrode recordings. The first algorithm uses optimal filters and successive source separation to maximize the signal to noise ratio and demix the single neuronal signals. The second approach shows that the filter output of the optimal matched filters can be interpreted in a Bayesian sense and can this way be used to derive a linear discriminant function based spike sorter. The performance of the methods is compared to that of others on simulated as well as unique real experimental data that is especially suited for that purpose and the real-time ability of the algorithms is discussed. However, the principal problem of benchmarking the quality of a spike sorting procedure on real data and the burdensome lack of a widely accepted benchmark pose a serious challenge. A way to overcome that obstacle with a community approach is proposed. A platform to host this approach is implemented in form of a website for the automatic and blind evaluation of spike sorting algorithms.