This title appears in the Scientific Report :
2023
A method for detecting spatio-temporal patterns of neuronal spikes based on principal component analysis
A method for detecting spatio-temporal patterns of neuronal spikes based on principal component analysis
Previous studies have suggested functional implications of spatio-temporal patterns (STPs) of neuronal spikes [1-3] with millisecond precision. Existing methods for STP detection are based on counting the occurrences of each STP found in the given data, and hence suffer from the combinatorial explos...
Saved in:
Personal Name(s): | Ito, Junji (Corresponding author) |
---|---|
Grün, Sonja | |
Contributing Institute: |
Computational and Systems Neuroscience; INM-6 Jara-Institut Brain structure-function relationships; INM-10 Computational and Systems Neuroscience; IAS-6 |
Imprint: |
2023
|
Conference: | Bernstein Conference 2023, Berlin (Germany), 2023-09-26 - 2023-09-29 |
Document Type: |
Poster |
Research Program: |
JL SMHB - Joint Lab Supercomputing and Modeling for the Human Brain (JL SMHB-2021-2027) Helmholtz Analytics Framework Human Brain Project Specific Grant Agreement 3 Human Brain Project Specific Grant Agreement 2 Neuroscientific Foundations |
Publikationsportal JuSER |
Previous studies have suggested functional implications of spatio-temporal patterns (STPs) of neuronal spikes [1-3] with millisecond precision. Existing methods for STP detection are based on counting the occurrences of each STP found in the given data, and hence suffer from the combinatorial explosion of the number of STPs to be considered, which can lead to massively multiple testing that deteriorates the detection power. Here we propose an alternative approach to STP detection, which is based on a principal component analysis (PCA) of spike train data and can circumvent the above-mentioned difficulties. Required parameters of the method are the width $\tau$ of the analysis time window (i.e., the maximum possible duration of STPs to be detected) and the allowed temporal imprecision $\sigma$ of the spike times within an STP. We first convolve parallel spike trains (of N neurons) with a Gaussian kernel of the standard deviation $\sigma$, thereby translating the spike trains into N parallel time series of instantaneous spike density functions. Next, we arbitrarily choose one of the N neurons, and for each spike of this neuron, we cut out a segment from the spike density time series starting at the spike time and ending at $\tau$ after it. We collect such segments (each of which is an N-by-$\tau f_s$ matrix; $f_s$: data sampling rate) for all the spikes of the chosen neuron (with M the number of spikes). We then concatenate rows of each of the segment matrices to reduce it into an N$\tau f_s$-dimensional row vector. Finally, we stuck these row vectors into an M-by-N$\tau f_s$ matrix, which we use as the data matrix for PCA. If the spike trains contain recurrences of an STP starting with a spike of the chosen neuron, the temporal modulations of spike density caused by the spikes in the STP are confined to specific columns of the data matrix, and hence the principal components (PCs) of this data matrix capture the correlated modulation of spike density along those columns. Application of the method to synthetic spike train data (10 neurons firing Poisson spikes at 10 Hz; 3 out of the 10 neuronsparticipating in an STP occurring at 0.5 Hz) successfully detects the STP as the first PC. We study the sensitivity and robustness of the proposed method by applying it to synthetic data with various parameter combinations and with temporally non-stationary and spatially inhomogeneous firing rates, and compare the performance to existing methods of STP detection. |