Cluster tendency assessment in neuronal spike data.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Data Analysis

Background:

• Spike sorting is essential for analyzing neuronal population activity from extracellular recordings.
• It is an unsupervised learning problem as the number of neurons is unknown.
• Existing methods require pre-specifying cluster numbers or post-clustering validation.

Purpose of the Study:

• To evaluate dimensionality reduction methods for improving spike sorting.
• To introduce and assess the improved visual assessment of cluster tendency (iVAT) for estimating cluster structures.
• To compare the reliability of clustering results obtained using different visualization and feature extraction techniques.

Main Methods:

• Investigated dimensionality reduction techniques for spike sorting visualization.
• Introduced and applied the improved visual assessment of cluster tendency (iVAT) method.
• Utilized t-Distributed Stochastic Neighbor Embedding (t-SNE) for data representation and feature extraction.
• Evaluated methods on datasets with ground truth labels, considering noise effects.

Main Results:

• iVAT effectively estimates cluster numbers for small datasets but can be less precise with larger numbers of clusters.
• Noise in extracellular recordings can negatively impact computational clustering.
• t-SNE provides superior data visualization, leading to more accurate spike sorting.
• Clusters derived from t-SNE features demonstrated higher reliability compared to other methods.

Conclusions:

• t-SNE is a valuable tool for both visualizing neuronal data and extracting features for enhanced spike sorting accuracy.
• Probabilistic clustering models are beneficial in handling noise during spike sorting.
• The findings suggest t-SNE can significantly improve the analysis of neuronal population activity.