paper Okun para problema de estabilidad.
Hong and Lieber. Novel electrode technologies for neural recordings. Nat Rev Neurosci. 2019 June
Chung et al. High-Density, Long-Lasting, and Multi-region
Electrophysiological Recordings Using Polymer
Electrode Arrays. Neuron. Jan 2019
DISTINGUIR CHRONIC DE SEMI CHRONIC
Fig. 1 shows raster plots from neurons recorded in consecutive days from the same microwire. The left panel shows responses to Anfield (Liverpool FC stadium). Although the responses seem to be stable across days, the waveforms associated to these neurons are very different. Therefore, without tracking the neural activity over the whole time between sessions, it would be hard to argue whether it is the same neuron or a different one from the network coding the concept “Anfield”. The right panel shows responses to Stonehenge on consecutive days. In this case, a sparse neuron in the first session leads to a very selective response. In the next session, the response seems to be maintained, but its selectivity is not the same. Tracking the activity would allow us to distinguish between a sorting error (the cluster associated to the neuron being contaminated by spikes from other neurons) and an actual change in selectivity.
[Gradual adaptation]
\cite{Calabrese_2011} Kalman filter to track the drifting of a fixed number of neurons
\cite{Pouzat_2004} Based on the construction of a Markov chain in the space of model parameters
and spike train configurations, where a configuration is defined by
specifying a neuron of origin for each spike. It cannot automatically find the “correct” number
of neurons in the data.
[tracking binario sigue o no, genericos sobre el sorter a usar]
\cite{Emondi_2004} The clusters with the most similar waveforms are assigned to the same neuron, provided their similarity exceeds a threshold. To set this threshold, we calculate two distributions: of within-file similarities, and of best matches in the across adjacent file similarities.
* \cite{Fraser_2012} Classifier using cross-correlograms, autocotrelogram, waveform and mean firing rate . It requires that the neurons should have some functional connectivity to display a rich cross-correlogram.
[ Bayesian Blocks]
\cite{Bar_Hillel_2006} Bayesian framework, with the source neurons modeled as a non-stationary mixture-of-Gaussians. At a first search stage, the data are divided into short time frames, and candidate descriptions of the data as mixtures-of-Gaussians are computed for each frame separately. At a second stage, transition probabilities
between candidate mixtures are computed, and a globally optimal clustering solution is found as the maximum-a-posteriori solution of the resulting
probabilistic model. It requires large samples of spikes with gaussian distribution.
High computational cost
\cite{Wolf_2009} A sequential extension of expectation maximization to fit a Gaussians mixture model (GMM) in each time frame using the clustering results from the previous time interval as a prior. The cluster means obtained in one time frame to seed the EM algorithm in the next time frame, while still allowing for signal non-stationarity and changing numbers of recorded neurons in transition between time frames.
Looks like can't handle sparse neurons, too few spikes per interval
\cite{Shalchyan_2014} Bayesian clustering approach that makes no assumptions on the distribution of the clusters and use kernel-based density estimation of the clusters in every time interval as a prior for Bayesian classification of the data in the subsequent time interval. The PCA basis of the first segment was fixed for extraction of features from spike waveforms of all time frames. It was also assumed that the spikes were classified into their neuronal sources in the first time segment.
*\cite{Shan_2017} Mixture of a fixed number of cluster with a t-distribution using windows of 1 minute. Has limitation with overclustering. The quality of the initial fit is an important factor in determining the user workload.
Loss of isolation occurs when two drifting clusters occupy the same region of feature space at different times, which appears as cluster overlap under a stationary analysis
\cite{Dhawale_2017} Sucesives SPC with all the detected spikes
[Spike sorting que soporta difting]
Mountainsort, support drift if the cluster doesn't have discontinuities and the clusters don't overlap in the feature spaces in any time \cite{Chung_2017}
Ironclus (new version of JRclus under development)
Kilosort
Combinato \cite{Niediek_2016} , using euclidean distances to join clusters
Methods
The core of the algorithms
Tracking Metric