opGMMassessment

Optimized Automated Gaussian Mixture Assessment

Overview

opGMMassessment provides optimized automated evaluation of the number and parameters of Gaussian mixtures in one-dimensional data. The package implements various methods for parameter estimation and for determining the number of modes in Gaussian mixture models (GMMs).

Motivation

Gaussian mixture models are probabilistic models commonly used in biomedical research to detect subgroup structures in data sets with one-dimensional information. Reliable model parameterization requires that the number of modes (i.e., states of the generating process) is known. However, this is rarely the case for empirically measured biomedical data.

Key Features

• Multiple algorithms for automatic GMM parameterization

• Various methods for mode number determination including:

  • Likelihood Ratio (LR) test
  • Akaike Information Criterion (AIC)
  • Bayesian Information Criterion (BIC)
  • Gap statistic (GAP)
  • Silhouette Index (SI)
  • NbClust methods
  • Figueiredo-Jain (FM) method

• Multiple GMM parameter estimation methods:

  • Markov Chain Monte Carlo (MCMC)
  • ClusterR GMM
  • Density-based Model-based Clustering (densityMclust)
  • Distribution Optimization (DO)
  • EM algorithm (normalmixEM)

• Parallel processing support for efficient computation

• Automatic calculation of Bayes decision boundaries

• Visualization with ggplot2

• Kolmogorov-Smirnov test for goodness of fit

Installation

Install from CRAN

install.packages("opGMMassessment")

Or install development version from GitHub

Newer version opGMMassessment_0.4.5

devtools::install_github("username/opGMMassessment")

Usage

library(opGMMassessment)

# Basic usage with default parameters (MCMC + Likelihood Ratio)
result <- opGMMassessment(Data = your_data)

# With custom parameters
result <- opGMMassessment(
  Data = your_data,
  FitAlg = "MCMC",           # Algorithm for GMM fitting
  Criterion = "LR",           # Criterion for mode number determination
  MaxModes = 8,               # Maximum number of modes to test
  MaxCores = 2,               # Number of cores for parallel processing
  PlotIt = TRUE,              # Display the plot
  KS = TRUE                   # Perform Kolmogorov-Smirnov test
)

# Access results
classes <- result$Cls          # Cluster assignments
means <- result$Means          # Means of Gaussian components
sds <- result$SDs              # Standard deviations
weights <- result$Weights      # Mixture weights
boundaries <- result$Boundaries # Bayes decision boundaries
plot <- result$Plot            # ggplot2 object
ks_test <- result$KS           # KS test results

Performance

Based on comparative evaluation on artificial and real-world data sets, the combination of the Likelihood Ratio (LR) test for mode number determination and Markov Chain Monte Carlo (MCMC) for parameter estimation consistently performs best and outperforms other available implementations.

Citation

If you use this package in your research, please cite:

Lötsch, J., Malkusch, S., & Ultsch, A. (2022). Comparative assessment of automated algorithms for the separation of one-dimensional Gaussian mixtures. Informatics in Medicine Unlocked, 34, 101113. https://doi.org/10.1016/j.imu.2022.101113

License

GPL-3

References

Lötsch, J., Malkusch, S., & Ultsch, A. (2022). Comparative assessment of automated algorithms for the separation of one-dimensional Gaussian mixtures. Informatics in Medicine Unlocked, 34, 101113.