Figure 5
Automatic K-means clustering and the cohesion-dispersion index (CD-index). (a) Some examples (one for each metric) considering different combinations (distance vs metric) for their comparative analysis using the proposed 24D feature vectors (Table 3). Each gray dotted square indicates the relationship between the value (in %) of the internal validation index (Silhouette, blue triangle; Davies-Bouldin, orange circle; or Dunn, green square) and the number of clusters obtained by the unsupervised method. Combinations (distance vs. metric) for which the Silhouette and Dunn indices reached their maximum values while the Davies-Bouldin index reached its minimum value are marked with an asterisk (*). In each of these cases, the suboptimal number of clusters is indicated (K = 3, 4 or 5). The distance-metric combinations for which the criterion above was not met are identified as failures. (b) CD-index value (in %) vs. number of clusters for the selected seven metrics. Note that the three internal validation indices interact to produce the maximal cohesion-dispersion of the clustering. Thus, the highest of all the CD-index values (i.e., CD = 100%, Cityblock vs. Correlation) afforded the optimal number of clusters (K = 3, marked with an asterisk over the red dashed rectangle). (c) Classification applying the SS-SPDF method/algorithm. Note that for the selected recording epoch (1.5 s; at rmPFC), two clusters were significant (cluster 1, 23 spikes; cluster 2, 19 spikes) in their configurations, while the third one was not (cluster 3, 1 outlier). (d) At the bottom right are illustrated the principal components analysis (PC1 vs. PC2 plot) and the waveform templates (magenta, brown, and gray profiles) for the resulting clusters. Note that, the action potentials clustering by the SS-SPDF method/algorithm (in c) are in perfect correspondence to those spikes (magenta circumferences, brown crosses, and gray square) clustering by principal component analysis (in d).
