Bayesian model comparison for compartmental models with applications in positron emission tomography

Yan Zhou, John A. D. Aston, Adam M. JohansenJournal of Applied Statistics 2013 · DOI

TL;DR

This paper introduces a robust Bayesian framework for analyzing Positron Emission Tomography (PET) data using compartmental models. It demonstrates that this approach offers more stable parameter estimation and better model selection than traditional methods, especially with noisy data, by incorporating biologically informed priors and a more flexible error model.

What Problem Does This Paper Address?

The analysis of Positron Emission Tomography (PET) data often relies on compartmental models to understand biochemical processes in the body, particularly in the brain. These models describe how tracer substances move between different compartments. However, analyzing PET data presents several challenges:

  • Scarcity of Data: PET scans typically yield a limited number of time points (20-30) for each region of interest (voxel).
  • High Noise Levels: The signal-to-noise ratio in PET data, especially at the voxel level, is often low.
  • Model Complexity: Identifying the correct number of compartments (which can correspond to different biological processes) is difficult.
  • Standard Methods Limitations: Traditional methods like Non-Negative Least Squares (NNLS) combined with information criteria (AIC, BIC) can be unstable and provide less reliable results, particularly with noisy data. They also struggle to quantify uncertainty effectively.

This paper aims to address these issues by developing a Bayesian approach that can simultaneously estimate parameters, compare different model structures (number of compartments), and quantify uncertainty, all while being robust to noise and incorporating prior biological knowledge.

Methods

The authors propose a Bayesian framework for compartmental model analysis that includes several key innovations:

  • Bayesian Model Comparison: Instead of relying solely on point estimates and information criteria, they use Bayesian model comparison to determine the most probable model structure (number of compartments) given the data. This involves calculating marginal likelihoods.
  • Markov Chain Monte Carlo (MCMC): To estimate the posterior distribution and marginal likelihoods, they employ a random-walk Metropolis algorithm, a standard MCMC technique.
  • Flexible Error Modeling: They demonstrate that the commonly used assumption of normally-distributed errors is inadequate for PET data. Instead, they propose using a more robust, heavier-tailed Student's t-distribution for the error structure, which better accounts for potential outliers and noise characteristics in PET data.
  • Biologically Informed Priors: The authors emphasize the importance of incorporating prior knowledge about the system's biophysics. They develop methods to construct informative priors for model parameters (like rate constants and volume of distribution) based on biological understanding, which helps to stabilize estimation and improve model selection, especially in noisy conditions.
  • Mass Univariate Approach: While acknowledging the potential for spatial modeling, they adopt a mass univariate approach, analyzing each voxel's time series independently. This is computationally intensive but allows for detailed analysis of individual regions.

Datasets

The paper uses two types of datasets:

  • Simulated Data: Data was generated from a three-compartment model with known parameters and varying levels of simulated noise to test the performance of their proposed methods against standard techniques.
  • Real PET Data: Two dynamic PET scans from a study using the tracer [11C]diprenorphine on normal subjects were analyzed. This data was previously used in other studies, allowing for direct comparison of results.

Key Insights and Ideas

  • Normal Error Assumption is Flawed: The assumption of normally-distributed errors, common in PET analysis, is shown to be insufficient and can lead to unstable estimates, especially with noisy data. A t-distribution provides a more robust alternative.
  • Bayesian Approach Outperforms Traditional Methods: The Bayesian framework, particularly with informative priors and a t-distributed error model, yields more stable parameter estimates (lower Mean Squared Error) and more reliable model selection compared to AIC, BIC, and NNLS, especially at higher noise levels.
  • Informative Priors Enhance Robustness: While vague priors provide a good starting point, incorporating biologically informed priors significantly improves the accuracy and stability of parameter estimation and model selection, particularly in challenging noisy scenarios.
  • Model Selection is Challenging but Informative: Even with the proposed Bayesian methods, reliably identifying the "true" model order can be difficult due to limited temporal data. However, the Bayesian approach provides posterior probabilities for each model, offering a measure of confidence in the chosen structure and revealing spatial patterns in model complexity across the brain.
  • Uncertainty Quantification is Crucial: The Bayesian framework naturally provides uncertainty estimates for parameters and model probabilities, which is invaluable for interpreting results and designing future studies, especially when comparing patient groups or different conditions.

Key Results

  • Parameter Estimation: For noisy PET data, the Bayesian approach with informative priors and a t-distributed error model achieved significantly lower Mean Squared Error (MSE) for the volume of distribution (VD) estimates compared to NLS and NNLS. For example, at a noise level of 5.12, the Bayesian informative prior method had an MSE of 0.195, while NLS was 1.191.
  • Model Selection: Bayesian model selection, especially with informative priors, showed improved performance in selecting the correct model order compared to AIC and BIC, particularly at higher noise levels. While not perfect, it provided a more nuanced view of model plausibility.
  • Computational Cost: The MCMC algorithm required approximately 0.15 seconds per voxel per core, leading to about two hours for a complete PET scan. While longer than NNLS, the authors argue this is acceptable given the increased information and accuracy.
  • Error Model Comparison: For typical PET data, the log marginal likelihood strongly favored the t-distributed error model over the normal error model, indicating its superior fit.

Limitations and Future Work

  • Computational Intensity: While the MCMC approach is efficient enough for current applications, it is computationally more intensive than traditional methods.
  • Spatial Homogeneity: The paper uses a mass univariate approach, analyzing each voxel independently. The authors suggest that future work could exploit spatial homogeneity within the brain to improve efficiency and parsimony.
  • Model Selection Reliability: Despite improvements, reliably identifying the true model order remains challenging due to data limitations. The authors advocate for cautious interpretation of model selection results and suggest model averaging as a more appropriate strategy when no single model is overwhelmingly preferred.
  • Further Algorithm Development: Work is ongoing to develop more sophisticated algorithms, potentially leveraging parallel hardware, to further accelerate the estimation procedure and explore alternative methods like path sampling.

Summary updated

Abstract

We develop strategies for Bayesian modelling as well as model comparison, averaging and selection for compartmental models with particular emphasis on those that occur in the analysis of positron emission tomography (PET) data. Both modelling and computational issues are considered. Biophysically inspired informative priors are developed for the problem at hand, and by comparison with default vague priors it is shown that the proposed modelling is not overly sensitive to prior specification. It is also shown that an additive normal error structure does not describe measured PET data well, despite being very widely used, and that within a simple Bayesian framework simultaneous parameter estimation and model comparison can be performed with a more general noise model. The proposed approach is compared with standard techniques using both simulated and real data. In addition to good, robust estimation performance, the proposed technique provides, automatically, a characterisation of the uncertainty in the resulting estimates which can be considerable in applications such as PET.
BibTeX
@article{zhou2013bayesian,
  author = {Yan Zhou and John A. D. Aston and Adam M. Johansen},
  title = {{Bayesian model comparison for compartmental models with applications in positron emission tomography}},
  journal = {Journal of Applied Statistics},
  year = {2013},
  doi = {10.1080/02664763.2013.772569},
}
Figure