A Study of Non-Linear Manifold Feature Extraction in Spike Sorting

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Advancements in neuronal recording hardware generate vast, complex datasets.
• Efficient processing requires capturing intrinsic neuronal activity relationships while mitigating noise.
• Automated spike sorting is crucial for analyzing electrophysiological data.

Purpose of the Study:

• To evaluate non-linear manifold feature extraction methods for automated spike sorting.
• To compare the efficacy of PHATE, t-SNE, UMAP, and TriMap against traditional methods like PCA.
• To identify the most adequate manifold learning technique for robust spike clustering.

Main Methods:

• Exploration of non-linear manifold feature extraction techniques (PHATE, t-SNE, UMAP, TriMap).
• Embedding high-dimensional spike shapes into low-dimensional manifolds.
• Clustering analysis of neuronal activity instances (spikes).
• Quantitative evaluation using clustering metrics (Adjusted Rand Index, Silhouette Score) on synthetic and real datasets.

Main Results:

• Non-linear manifold methods produced more separable and robust spike clusters compared to PCA.
• Several manifold feature extraction techniques demonstrated superior performance.
• The study utilized 95 synthetic and 2 real single-channel datasets.

Conclusions:

• Non-linear manifold embeddings offer a high-precision approach for next-generation electrophysiological spike sorting.
• These methods enhance the clarity and reliability of neuronal data analysis.
• Future work should explore multi-channel data and advanced manifold techniques.