Figure 3
GMM-based spike sorting framework. (A) Example of feature values (wavelet coefficients) and their GMM fittings. In the proposed framework, wavelet coefficients (or PC scores) are ranked according to a clustering separability metric; 4 metrics were investigated: var, Ipeak, Iinf or Idist (see Materials and Methods). The example depicts wavelet coefficient distributions ranked by the Idist metric. Note that unimodal distributions have lower Idist values. (B) The first two wPCs of the coefficients in A. Weighted PCA was obtained by normalizing the variance of wavelet coefficients by Idist and applying PCA. Clustering was done using the first 5 wPCs. For pure PCA and WD, we used the first 5 features according to the separability metric. (C) The probability density function of the 5-dimension GMM computed from the feature subspace (same dimensions as in B) and its peaks (white dots). The GMM was computed with 12 (Dataset A) or 20 (Dataset B) Gaussians. (D) Representation of a fixed-mean GMM with Gaussians centered at the peaks in C. Dots and ellipses denote the center and the 2-standard-deviation boundary of each Gaussian; line thickness represents Gaussian amplitude. (E) Final classification of the waveforms using the fixed-mean GMM. Each waveform was assigned to the Gaussian with higher probability in the model in D. Colors in D and E represent different Gaussians/clusters. For this example, the spikes of three neurons were analyzed. Notice that the algorithm performs “overclustering”, capturing changes in the waveform shape of a same neuron due to bursting activity. Clusters can be merged during post-processing adjustments.
