We used different metrics to quantify the performance. Recall, precision and accuracy were calculated in the same way as defined by Jun et al. \cite{Jun_2017}. Given the set of spikes from a simulated neuron k, and a class c from a given algorithm (i.e. a putative single neuron), \(n^{k,c}_{match}\) is defined as the number of spikes that belong to the intersection of both sets, \(n^{k,c}_{fp}\) is defined as the number of spikes belonging to class c that are not from the neuron k, and \(n^{k,c}_{miss}\) is defined as the number of spikes from neuron k that are not included in class c. Then, the metrics are defined as:
\(precision\left(k,c\right)\ =\ \frac{n^{k,c}_{match}}{n^{k,c}_{match}+n^{k,c}_{fp}}\)
\(recall\left(k,c\right)\ =\ \frac{n^{k,c}_{match}}{n^{k,c}_{miss}+n^{k,c}_{match}}\)
\(accuracy\left(k,c\right)\ =\ \frac{n^{k,c}_{match}}{n^{k,c}_{match}+n^{k,c}_{miss}+n^{k,c}_{fp}}\)
Finally, for each neuron k the class c* with the highest accuracy is used as the best match, and the metrics for the pair (k,c*) are the ones reported for that neuron.
Figure \ref{582093} shows a simulated long-term recording with 4 neurons (top) and the results obtained by each algorithm. The borders of each block calculated by Spikes_Link are shown with dotted lines, and the mean waveform of the classes within each block was computed. Spikes_Link_WC was able to track the classes with high performance (\(\mu_a\)=0.998, \(\mu_p\)=0.99, and \(\mu_r\)=0.99, for accuracy, precision, and recall), but the other algorithms (particularly the standard Wave_clus) tend to split the simulated neurons in multiple classes (overclustering). In this simulation, Wave_clus did not merge parts of the simulated neurons (as reflected by its precision with \(\mu_p\)=0.99), but Combinato merged together two simulated neurons (as reflected by its precision with \(\mu_p\)=0.75). This is due to the final step of Combinato, which merges together all the classes with mean waveforms that are close enough in euclidean space without using temporal information.