A Study of Deep Clustering in Spike Sorting

Benchmarking Deep Clustering Algorithms for Next-Generation Spike Sorting

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

"A Study of Deep Clustering in Spike Sorting" by Eugen-Richard Ardelean and Raluca Laura Portase, published in Neuroinformatics, 2025.

---

Overview

This project provides a large-scale benchmark of 12 deep clustering algorithms against the traditional spike sorting pipeline (feature extraction combined with K-means clustering).

Traditional spike sorting separates representation learning and clustering into distinct steps, which may not optimally capture the complex structure of spike data. Deep clustering addresses this by performing a dual optimization, effectively learning non-linear representations tailored for clustering. Our findings indicate that these deep clustering approaches are the most suitable methods for accurately identifying individual neuronal activity in modern multi-electrode recordings.

---

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.

---

Methods Benchmarked

The study compares deep clustering against the traditional two-stage pipeline.

Deep Clustering Algorithms

A total of 12 deep clustering algorithms were benchmarked.

• Key Algorithms: ACeDeC, AEC, DCN, DDC, DEC, DeepECT, DipDECK, DipEncoder, DKM, IDEC, VaDE, N2D.

Traditional Pipeline

• Feature Extraction: Linear (PCA, ICA) and Non-linear/Manifold (Isomap, LLE, t-SNE, Diffusion Map) methods.
• Clustering: Clustering was performed using K-Means for evaluation of the feature extraction methods.

---

Results Summary

Performance was evaluated using six 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:

• Superior Performance: A subset of deep clustering algorithms—particularly ACeDeC, DDC, DEC, IDEC, and VaDE—significantly outperformed traditional methods, especially as dataset complexity increased.
• Top Performers:
  • DDC excelled on datasets with low to medium cluster counts.
  • DEC, IDEC, and VaDE were the top performers for datasets with a medium to high number of clusters.

---

Citation

If you use this work, please cite:

@article{Ardelean2025DeepClustering,
  title     = {A Study of Deep Clustering in Spike Sorting},
  author    = {Ardelean, Eugen-Richard and Portase, Raluca Laura},
  journal   = {Neuroinformatics},
  year      = {2025},
  volume    = {23},
  pages     = {51},
  doi       = {10.1007/s12021-025-09751-4},
}

---

📬 Contact

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

---