Comparison of Uniform Manifold Approximation and Projection (UMAP)-based and feature-based spike sorting on synthetic data.
<p><b>(A)</b> In a simulated dataset containing three different neurons, UMAP-based sorting forms three well-separated clusters with distinct shapes and colors, indicating strong accuracy and minimal overlap. <b>(B)</b> Applying wavelet sorting to the same dataset merge...
Guardat en:
| Autor principal: | |
|---|---|
| Altres autors: | , , , , , , , , , , |
| Publicat: |
2025
|
| Matèries: | |
| Etiquetes: |
Afegir etiqueta
Sense etiquetes, Sigues el primer a etiquetar aquest registre!
|
| Sumari: | <p><b>(A)</b> In a simulated dataset containing three different neurons, UMAP-based sorting forms three well-separated clusters with distinct shapes and colors, indicating strong accuracy and minimal overlap. <b>(B)</b> Applying wavelet sorting to the same dataset merges the spikes into just two clusters that exhibit noticeable overlap and weak separation, as illustrated in both 3D (top) and 2D (bottom) views. <b>(C)</b> Principal component analysis (PCA)-based sorting isolates only a single cluster with poor separation, underscoring PCA’s difficulty with complex spike distributions. <b>(D)</b> By contrast, UMAP spike sorting consistently surpasses both wavelet and PCA methods, particularly in higher-dimensional spaces, as measured by the F1 score. <b>(E)</b> UMAP-based sorting also proves robust against noise perturbations at various spike dilution levels (dark trace; firing rate [fr] at 100% vs. light blue trace; fr diluted at 40%). Even under increased noise and cluster dilution, it preserves high F1 scores. <b>(F)</b> Moreover, UMAP-based sorting (dark blue) remains accurate despite data loss, achieving high F1 scores even when a large fraction of spikes (x-axis) is removed from one cluster, whereas wavelet (green) and PCA (orange) methods experience a significant decline in performance. The synthetic data used to generate this figure are publicly available at [<a href="http://www.plosbiology.org/article/info:doi/10.1371/journal.pbio.3003527#pbio.3003527.ref037" target="_blank">37</a>], and the code for performing the analyses is available at [<a href="http://www.plosbiology.org/article/info:doi/10.1371/journal.pbio.3003527#pbio.3003527.ref049" target="_blank">49</a>].</p> |
|---|