Clustering action potential spikes: Insights on the use of overfitted finite mixture models and Dirichlet process mixture models

Zoé van Havre, Nicole White, Judith Rousseau, Kerrie MengersenarXiv 2016 · DOI · arXiv

Abstract

The modelling of action potentials from extracellular recordings, or spike sorting, is a rich area of neuroscience research in which latent variable models are often used. Two such models, Overfitted Finite Mixture models (OFMs) and Dirichlet Process Mixture models (DPMs) are considered to provide insights for unsupervised clustering of complex, multivariate medical data when the number of clusters is unknown. OFM and DPM are structured in a similar hierarchical fashion but they are based on different philosophies with different underlying assumptions. This study investigates how these differences impact on a real study of spike sorting, for the estimation of multivariate Gaussian location-scale mixture models in the presence of common difficulties arising from complex medical data. The results provide insights allowing the future analyst to choose an approach suited to the situation and goal of the research problem at hand.

Introduction

In neuroscience, monitoring neural activity through extracellular recordings presents a significant technical hurdle known as "spike sorting." This task requires researchers to take a raw stream of electrical pulses (action potentials) and determine which specific neuron produced each pulse. Fundamentally, this is an unsupervised clustering problem where the number of sources—the neurons—is unknown. Standard clustering methods often struggle with the inherent complexities of medical data, such as outliers, overlapping signal profiles, and uncertainty regarding the true number of clusters. This paper evaluates two Bayesian philosophies for addressing these challenges: Overfitted Finite Mixture (OFM) models and Dirichlet Process Mixture (DPM) models.

Methodology

The researchers compared two specific Bayesian latent variable models to handle cluster uncertainty:

  1. Overfitted Finite Mixture (OFM): This approach starts by assuming a maximum number of clusters ($K^*$) that is intentionally larger than the expected true number. By employing a Dirichlet prior with very small hyperparameters, the model encourages "posterior emptying," effectively forcing unnecessary clusters to have zero weight. A Prior Parallel Tempering MCMC algorithm is used to navigate the resulting posterior landscape.
  2. Dirichlet Process Mixture (DPM): Often referred to as an "infinite mixture model," the DPM does not assume a fixed number of clusters. Instead, it uses a Dirichlet Process prior that allows the number of clusters to grow dynamically with the data. This was implemented using a slice sampler to avoid manual truncation.

To mitigate the influence of outliers, both models were applied to data reduced to four dimensions using Robust Principal Component Analysis (PCA).

Data Description

The study utilized real-world clinical data consisting of extracellular recordings from the subthalamic nucleus of patients with advanced Parkinson’s Disease. The data was collected during Deep Brain Stimulation (DBS) surgery across three independent datasets ($Y_1, Y_2, Y_3$) with sample sizes ranging from 192 to 348 spikes.

Original action potential spikes for datasets Y1, Y2, and Y3, demonstrating the complexity and overlap of the neural signals.

Results and Discussion

The most significant finding is the contrast between "lumping" and "splitting" behaviors. The OFM tends to "lump" spikes that do not fit well into a single, high-variance "noise" cluster. This provides a sparse, high-confidence summary of the probable set of models. In contrast, the DPM acts as a "splitter," creating a "long tail" of many tiny clusters to explain every data point, thus providing information about all possible models.

For well-defined, high-density clusters, both models showed high agreement. However, differences emerged at the data's fringes. In dataset $Y_3$—the most complex—the OFM identified four main groups with high probability. The DPM identified one large group but fragmented the low-density regions into seven smaller "tail" groups.

Histogram of the third principal component (PC3) for dataset Y3, illustrating the feature distribution used for clustering.

Conclusion

While both Bayesian approaches effectively group neural activity, the OFM is superior for generating sparse models that handle noise gracefully. The DPM offers a more complex, exploratory view that can capture fine structural details but may lead to over-complication in noisy environments. Future research could explore non-Gaussian kernels to better fit skewed spike distributions and incorporate time-dependent components to account for non-stationary neural signals.

Summary updated

Abstract

The modelling of action potentials from extracellular recordings, or spike sorting, is a rich area of neuroscience research in which latent variable models are often used. Two such models, Overfitted Finite Mixture models (OFMs) and Dirichlet Process Mixture models (DPMs) are considered to provide insights for unsupervised clustering of complex, multivariate medical data when the number of clusters is unknown. OFM and DPM are structured in a similar hierarchical fashion but they are based on different philosophies with different underlying assumptions. This study investigates how these differences impact on a real study of spike sorting, for the estimation of multivariate Gaussian location-scale mixture models in the presence of common difficulties arising from complex medical data. The results provide insights allowing the future analyst to choose an approach suited to the situation and goal of the research problem at hand.
BibTeX
@misc{havre2016clustering,
  author = {Zoé van Havre and Nicole White and Judith Rousseau and Kerrie Mengersen},
  title = {{Clustering action potential spikes: Insights on the use of overfitted finite mixture models and Dirichlet process mixture models}},
  year = {2016},
  eprint = {1602.01915},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1602.01915},
}
Figure