Using Bayesian latent Gaussian graphical models to infer symptom\n associations in verbal autopsies

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

TL;DR

This paper introduces a Bayesian latent Gaussian graphical model to improve cause-of-death assignments in "verbal autopsies." By moving beyond the simplistic assumption that symptoms are independent, the model uses expert knowledge and sparse data to map complex symptom associations, leading to significantly more accurate public health data.

What Problem Does This Paper Address?

In many parts of the world, deaths go unrecorded by medical professionals. To understand why people are dying, researchers use Verbal Autopsies (VAs)—structured interviews with the family or caregivers of the deceased. These surveys result in a messy mix of binary (Yes/No), continuous (How many days?), and count data.

The massive bottleneck is how to turn these symptom reports into a "Cause of Death" (CoD). Most current algorithms rely on a "Naive Bayes" assumption: they assume that, given a cause like Malaria, having a fever is entirely independent of having a cough. In reality, symptoms are deeply interconnected. Ignoring these correlations leads to biased results, especially when data is sparse or full of missing values. The challenge is learning these complex "webs" of symptoms when you don't have the "copious, high-quality data" that standard statistical methods require.

Methods

The authors propose a Bayesian latent Gaussian graphical model tailored for mixed-type data. Here’s the breakdown of their approach:

  • Latent Gaussian Representation: Since symptoms come in different formats (binary, continuous, etc.), the model maps them all to a "latent" (hidden) multivariate Gaussian space. This allows the researchers to model dependencies on a unified, smooth scale.
  • Spike-and-Slab Priors: To avoid "overfitting" or finding fake associations, they use a "spike-and-slab" prior on the inverse correlation matrix. Think of the "spike" as a way to force weak connections to zero (simplifying the graph) and the "slab" as a way to allow strong, meaningful associations to stand out.
  • Informative Priors: Unlike previous models that treat marginal distributions as a nuisance, this model incorporates expert clinical knowledge about how common a symptom is for a specific disease. This "expert boost" helps the model learn the structure even when training data is limited.
  • Gaussian Mixture Framework: They extend the model into a mixture framework, allowing it to simultaneously cluster deaths into causes and learn the unique symptom patterns of the population.

Datasets

The model was evaluated using three primary sources:

  1. Simulated Data: Used to benchmark the model against known "ground truths" under various levels of missing data (up to 50%).
  2. PHMRC Gold Standard Dataset: A massive dataset of ~7,000 deaths from six global sites where the cause of death was confirmed by hospital pathology/medical imaging.
  3. Karonga HDSS Data: Real-world data from a surveillance site in Malawi containing ~1,900 adult deaths, where causes were assigned by local physicians.

Key Insights and Ideas

  • The Cost of Independence: The authors demonstrate that the standard "independence assumption" isn't just a simplification—it’s a source of structural bias. It often leads models to over-assign deaths to "undetermined" or "unspecified" categories because it can't recognize the specific combinations of symptoms that define a disease.
  • Expert Knowledge as a Scaffold: One of the most powerful findings is that even if expert priors are slightly "off" (misspecified), they still provide enough of a scaffold for the model to learn the underlying dependence structure better than a model with no prior knowledge at all.
  • Questionnaire Optimization: By identifying which symptoms are highly correlated (and thus redundant), this model can inform the design of shorter, more efficient VA questionnaires, reducing the "survey fatigue" for grieving families.

Key Results

  • Superior Accuracy: In the PHMRC cross-validation study, the proposed Gaussian Mixture model consistently outperformed established methods like Tariff, InterVA, and InSilicoVA in both CSMF (Cause-Specific Mortality Fraction) accuracy and "Top-1" individual cause assignment.
  • Robustness to Missingness: Even when 50% of the data was missing, the spike-and-slab GGM maintained high performance in recovering the true correlation structure, whereas semi-parametric alternatives degraded quickly.
  • Clinical Discovery: In the Malawi data, the model successfully recovered known clinical associations, such as the strong positive link between a history of high blood pressure and paralysis on one side of the body (a classic indicator of a stroke).

Limitations and Future Work

While powerful, the model relies on Markov chain Monte Carlo (MCMC) sampling, which can be computationally expensive and prone to getting "stuck" in local modes when dealing with extreme clinical values (probabilities near 0 or 1).

The authors suggest that future work could involve:

  • Deterministic Algorithms: Moving toward EM-type algorithms to speed up computation.
  • Group-Specific Correlations: Allowing different causes of death to have their own unique symptom correlation matrices (e.g., the way symptoms cluster for TB might be structurally different from how they cluster for Heart Disease).
  • Ordinal Data: Expanding the model to handle "ranked" symptoms (e.g., mild, moderate, severe) more naturally.

Summary updated

Abstract

Learning dependence relationships among variables of mixed types provides\ninsights in a variety of scientific settings and is a well-studied problem in\nstatistics. Existing methods, however, typically rely on copious, high quality\ndata to accurately learn associations. In this paper, we develop a method for\nscientific settings where learning dependence structure is essential, but data\nare sparse and have a high fraction of missing values. Specifically, our work\nis motivated by survey-based cause of death assessments known as verbal\nautopsies (VAs). We propose a Bayesian approach to characterize dependence\nrelationships using a latent Gaussian graphical model that incorporates\ninformative priors on the marginal distributions of the variables. We\ndemonstrate such information can improve estimation of the dependence\nstructure, especially in settings with little training data. We show that our\nmethod can be integrated into existing probabilistic cause-of-death assignment\nalgorithms and improves model performance while recovering dependence patterns\nbetween symptoms that can inform efficient questionnaire design in future data\ncollection.\n
BibTeX
@inproceedings{li2019using,
  author = {Zehang Li and Tyler H. McCormick and Samuel J. Clark},
  title = {{Using Bayesian latent Gaussian graphical models to infer symptom\textbackslash\{\}n associations in verbal autopsies}},
  booktitle = {Bayesian},
  year = {2019},
  doi = {10.1214/19-ba1172},
}
Figure