Learning to Address Health Inequality in the United States with a Bayesian Decision Network

Tavpritesh Sethi, Anant Mittal, Shubham Maheshwari, Samarth ChughAAAI 2019 · DOI

Introduction

Health inequality in the United States represents a significant public health crisis, with an estimated annual economic burden of $300 billion. While factors such as socioeconomic status, race, and geography are known to influence life expectancy, they are frequently studied in isolation. This siloed approach hinders the ability of policymakers to understand the complex, interconnected drivers of the "longevity gap"—the disparity in life expectancy between the wealthiest and poorest quartiles of the population.

This research proposes a data-driven Bayesian Decision Network (BDN) to model these dependencies. By moving beyond simple predictive modeling toward interpretable and explainable AI, the authors provide a transparent framework for identifying high-payoff policy interventions that can effectively narrow the longevity gap.

Methodology

The study utilizes a massive, granular dataset from the Health Inequity Project, integrating 1.4 billion tax records, county-level mortality data, healthcare indices from the Dartmouth Atlas, and health behavior data from the CDC. The final merged dataset covers 1,559 counties and includes features ranging from education levels to annual lipid testing rates.

The researchers employed a multi-step probabilistic graphical modeling approach:

  1. Structure Learning: A hill-climbing algorithm combined with the Bayesian Information Criterion (BIC) was used to identify dependencies. To ensure robustness, ensemble averaging was applied across 1001 bootstrap samples, retaining only the most consistent edges.
  2. Causal Constraints: Geographic variables (State, County) were constrained to act as drivers rather than consequences, facilitating policy reasoning.
  3. Bayesian Decision Networks (BDNs): The standard Bayesian Network was extended with Utility nodes (representing desired outcomes like longevity) and Decision nodes (representing actionable policy interventions).
  4. Inference: Both exact and approximate inference methods were used to calculate the probability of health outcomes based on specific policy choices.
The learned Bayesian Network structure, highlighting key clusters of influence such as diversity (Q1), preventive care (Q2), and education (O3).

Key Findings

The learned network revealed several critical drivers of health outcomes and the longevity gap:

  • Diversity as a Protective Factor: Population diversity (specifically the proportion of foreign-born residents) emerged as a primary "parent node" of the longevity gap. In low-diversity counties, the probability of a longevity gap exceeding 9 years was 41%, whereas in high-diversity counties, this probability plummeted to just 3%.
  • The Power of Preventive Care: The "Index of Preventive Care" was identified as a direct neighbor of life expectancy. Improving the quality of preventive care to the highest stratum significantly shifted the probability distribution of life expectancy toward older ages.
  • Education as a Central Hub: Graduate-level education acted as a major distributor of influence in the network, directly impacting obesity rates, exercise habits, income levels, and unemployment.
  • Social Fabric: The model confirmed strong links between low social mobility, family instability (e.g., children born to single mothers), and income inequality (Gini index).
Conditional probability of female life expectancy (le_mean_pool_F) based on the quality of preventive care (med_prev_qual_z). Higher quality care (top) significantly increases the probability of living beyond age 85 compared to lower quality care (bottom).

Policy Implications and Results

The BDN framework allows for the simulation of specific interventions to determine their potential public health payoff. For instance, the model identified Annual Lipid Testing for diabetic populations as the specific preventive measure with the highest individual payoff for maximizing longevity. Furthermore, the results suggest that improving preventive care quality can increase the probability of living beyond age 85 by 43% for females and 30% for males.

Conclusion

While the study relies on observational data, the use of structural constraints and ensemble averaging provides a robust foundation for causal reasoning. By identifying actionable "parent" nodes like preventive care and social diversity, this Bayesian approach offers a roadmap for targeted public health interventions. To facilitate real-world application, the authors developed an interactive web application for policymakers to visualize these dependencies and simulate the impacts of various social and medical interventions in real-time.

Summary updated

BibTeX
@inproceedings{sethi2019learning,
  author = {Tavpritesh Sethi and Anant Mittal and Shubham Maheshwari and Samarth Chugh},
  title = {{Learning to Address Health Inequality in the United States with a Bayesian Decision Network}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.3301710},
}
Figure