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

Benchmarking Manifold Learning for Next-Generation Neural Data Analysis

This repository contains the code and configurations used in the study:

"A Study of Non-Linear Manifold Feature Extraction in Spike Sorting"
by Eugen-Richard Ardelean and Raluca Portase, published in Neuroinformatics, 2025.

---

Overview

This project investigates non-linear manifold learning methods for spike sorting, a key step in processing extracellular neuronal recordings. We compare modern approaches such as PHATE, UMAP, TriMap, t-SNE, Diffusion Maps, Isomap, LLE against classical methods (PCA, ICA, MDS, Autoencoders) to determine which feature extraction techniques produce the most separable and robust embeddings of spike waveforms.

Experiments were conducted on:

• 95 synthetic datasets (with ground truth cluster labels).
• 2 real datasets (spe-1 recordings, Marques-Smith et al.) with dual extracellular/intracellular ground truth validation.

Results show that UMAP, PHATE, and TriMap consistently outperform other methods, offering strong separability, robustness to noise, and scalability for high-density probes.

---

Datasets

Synthetic Spike Waveforms (Pedreira et al., 2012)

• Description: 95 single-channel synthetic datasets derived from real monkey recordings.
• Characteristics: 2–20 clusters per dataset, ~9,300 spikes on average, including multi-unit clusters.
• Usage: Benchmarking feature extraction across diverse conditions.
• Access: Publicly available.

Real Datasets (spe-1, Marques-Smith et al., 2018/2020)

• Description: Patch-clamp + 384-channel CMOS extracellular recordings in rat cortex.
• Ground Truth: Dual intracellular/extracellular data for 21 neurons.
• Datasets Used: c28 and c37.

---

Feature Extraction Methods

• Linear: PCA, MDS, ICA
• Non-linear: KPCA, SOM, Autoencoders
• Manifold learning: LLE, MLLE, HLLE, LTSA, Isomap, Spectral Embedding, Diffusion Maps, PHATE, UMAP, TriMap, t-SNE

Clustering was performed using K-Means for evaluation.

---

Results Summary

Performance was evaluated using clustering metrics: Adjusted Rand Index (ARI), Adjusted Mutual Information (AMI), Purity, Silhouette Score (SS), Calinski-Harabasz Score (CHS), and Davies-Bouldin Score (DBS).

Key findings:

• UMAP, PHATE, TriMap: best overall performance across metrics and datasets.
• t-SNE, Autoencoders: good separability but slower runtime.
• Diffusion Maps, LLE, MLLE: strong only for datasets with few clusters, degrade with complexity.
• PCA, ICA, MDS, SOM: consistently underperformed.

---

Citation

If you use this work, please cite:

@article{Ardelean2025Manifold,
  title     = {A Study of Non-Linear Manifold Feature Extraction in Spike Sorting},
  author    = {Ardelean, Eugen-Richard and Portase, Raluca},
  journal   = {Neuroinformatics},
  year      = {2025},
  volume    = {23},
  pages     = {48},
  doi       = {10.1007/s12021-025-09744-3}
}

---

📬 Contact

For questions, please contact: 📧 ardeleaneugenrichard@gmail.com

---