Using Bayesian Latent Gaussian Graphical Models to Infer Symptom Associations in Verbal Autopsies

Zehang Li, Tyler H. McCormick, Samuel J. Clark

TL;DR

This paper introduces a novel Bayesian approach using latent Gaussian graphical models to uncover symptom associations in verbal autopsy (VA) data, which is often sparse and incomplete. This method improves the accuracy of cause-of-death assignment and provides a more nuanced understanding of symptom relationships, crucial for public health in resource-limited settings.

What Problem Does This Paper Address?

In many parts of the world, accurately determining the cause of death is a significant challenge due to the lack of systematic medical autopsies. Verbal autopsies (VAs) are used as a proxy, involving interviews with caregivers about symptoms and circumstances surrounding a death. However, VA data is typically sparse, contains many missing values, and consists of mixed data types (binary, continuous, count). Existing statistical methods often assume symptom independence, which can lead to inaccurate cause-of-death assignments. Furthermore, these methods typically require large, high-quality training datasets, which are rarely available in the low-resource settings where VAs are most needed. This paper aims to bridge this gap by developing a method that can learn complex symptom associations from limited, messy data and improve cause-of-death classification.

Methods

The core of the paper is a Bayesian latent Gaussian graphical model. This model assumes that the observed mixed-type data (VA symptoms) can be represented by underlying latent Gaussian variables. The dependence structure between these latent variables is modeled using a graphical model, where edges represent associations.

Key methodological contributions include:

  • Latent Gaussian Graphical Model: This framework allows for modeling complex dependencies between variables of mixed types, moving beyond the simplifying independence assumptions of many existing methods.
  • Incorporation of Informative Priors: The authors emphasize the use of prior knowledge about the marginal distributions of symptoms. This is crucial when data is sparse, as it helps to guide the model and improve the estimation of the dependence structure.
  • Spike-and-Slab Prior: For inferring sparse inverse correlation matrices, they propose a spike-and-slab prior. This prior is designed to simultaneously perform graph selection (identifying which symptom associations are important) and parameter inference, making it computationally efficient and effective for identifying key relationships.
  • Integration with Cause-of-Death Assignment: The latent Gaussian graphical model is extended to a Gaussian mixture model framework. This allows it to be seamlessly integrated into existing probabilistic cause-of-death assignment algorithms, improving their performance.
  • Efficient Sampling Algorithms: The paper details Markov Chain Monte Carlo (MCMC) algorithms, including block updates and elliptical slice sampling (ESS), to efficiently sample from the posterior distributions of the model parameters.

Datasets

The paper utilizes two primary datasets for evaluation:

  • PHMRC Gold Standard VA Dataset: This dataset, comprising approximately 7,000 deaths from six sites across four countries, serves as a "gold standard" for training and validation. It was used to evaluate the performance of correlation matrix estimation, graph recovery, and classification accuracy.
  • Karonga Health and Demographic Surveillance System (HDSS) Data: This dataset from Malawi, with a higher proportion of missing data and smaller sample sizes, is used to evaluate the method in a more realistic, challenging scenario.

Key Insights and Ideas

  • Prior Knowledge is Powerful: The study strongly demonstrates that incorporating prior information about symptom marginal distributions can significantly improve the estimation of dependence structures, especially when data is scarce. This is a critical insight for applying statistical models in data-poor environments.
  • Symptom Associations Matter: Ignoring symptom dependencies, as many traditional VA methods do, can lead to biased cause-of-death assignments. Uncovering these associations provides a more accurate picture of disease patterns.
  • Symptom Relationships Inform Design: The identified symptom associations can guide the design of future VA questionnaires, making them more efficient and informative.
  • Flexibility for Mixed Data: The latent Gaussian graphical model provides a flexible framework for handling the mixed-type nature of VA data, avoiding the need for potentially lossy pre-processing steps like dichotomization.

Key Results

  • Improved Classification Accuracy: The proposed Bayesian latent Gaussian graphical model consistently outperformed existing methods (e.g., Tariff, InterVA, Naive Bayes, InSilicoVA) in terms of cause-of-death assignment accuracy (CSMF accuracy, Top 1, Top 2, Top 3 cause accuracy) across various simulation scenarios and real-world datasets. For instance, in simulations, the proposed method achieved better AUC and F1 scores, especially with high missing data.
  • Robustness to Missing Data and Misspecification: The model showed robustness to missing data and misspecified prior information, a significant advantage in real-world VA applications.
  • Accurate Correlation Matrix Estimation: The method demonstrated superior performance in recovering the true correlation matrix compared to semi-parametric and other Bayesian graphical models, particularly under missing data conditions.
  • Identification of Key Symptom Dependencies: The analysis of the Karonga data revealed interesting symptom pairs with conditionally dependent latent factors, such as the association between high blood pressure history and paralysis, which aligns with medical understanding.

Limitations and Future Work

  • Computational Cost: While efficient MCMC algorithms are employed, estimating complex mixture models can still be computationally intensive, potentially suffering from slow mixing when the sampler gets trapped in local modes, especially with strong prior information.
  • Interpretation of Latent Variables: The learned dependencies are between latent variables, not directly between observed binary symptoms. While this provides insights into underlying processes, direct interpretation of these latent relationships might require further work.
  • Extension to Ordinal Data: The current model primarily handles binary and continuous data. Extensions to ordinal data are suggested.
  • Group-Specific Correlation Matrices: The paper assumes a single correlation matrix across all mixtures. Extending this to group-specific correlation matrices, though challenging due to unbalanced mixture probabilities, is a potential future direction.
  • Symptom Selection: The authors suggest that the framework could be extended for data-driven symptom selection, aiming to reduce questionnaire length and cost.
  • Alternative Inference Methods: For computational efficiency, deterministic EM-type algorithms are proposed as an alternative to MCMC for targeting posterior modes.

Summary updated

Abstract

Learning dependence relationships among variables of mixed types provides insights in a variety of scientific settings and is a well-studied problem in statistics. Existing methods, however, typically rely on copious, high quality data to accurately learn associations. In this paper, we develop a method for scientific settings where learning dependence structure is essential, but data are sparse and have a high fraction of missing values. Specifically, our work is motivated by survey-based cause of death assessments known as verbal autopsies (VAs). We propose a Bayesian approach to characterize dependence relationships using a latent Gaussian graphical model that incorporates informative priors on the marginal distributions of the variables. We demonstrate such information can improve estimation of the dependence structure, especially in settings with little training data. We show that our method can be integrated into existing probabilistic cause-of-death assignment algorithms and improves model performance while recovering dependence patterns between symptoms that can inform efficient questionnaire design in future data collection.
BibTeX
@inproceedings{li2019using,
  author = {Zehang Li and Tyler H. McCormick and Samuel J. Clark},
  title = {{Using Bayesian Latent Gaussian Graphical Models to Infer Symptom Associations in Verbal Autopsies}},
  booktitle = {Bayesian},
  year = {2019},
  doi = {10.1214/19-ba1172},
}

Loading related content...

Figure