Global variation in cardiometabolic risk structures: A 48-country comparative Bayesian network analysis in 146,000 participants using WHO STEPS data

This study analyzed 146,000 participants across 48 countries using Bayesian networks to reveal significant regional variations in cardiometabolic risk structures, demonstrating that public health strategies must be tailored to specific geographic contexts rather than relying on generalized models.

The Gist

Imagine that heart disease and diabetes (what the paper calls "cardiometabolic diseases") aren't caused by a single switch being flipped, but by a complex, tangled web of strings. Some strings are things you can't change, like your age or where you were born. Others are things you can change, like what you eat, how much you move, or whether you smoke.

This paper is like a team of detectives trying to map out exactly how these strings are tied together in 48 different countries. Instead of just looking at one string at a time, they used a special computer tool called a Bayesian Network to see the whole web at once. Think of this tool as a "relationship map" that shows which factors pull on which other factors.

Here is what they found, explained simply:

1. The Map Looks Different in Every Neighborhood

The researchers discovered that the "web of risk" looks different depending on where you are in the world.

• The Analogy: Imagine trying to understand why traffic jams happen. In one city, the jam might be caused mostly by a bad intersection (a specific road layout). In another city, the jam might be caused by too many people leaving work at the same time. Even though the result (a traffic jam) is the same, the cause and the connections between causes are totally different.
• The Finding: The computer maps for countries in the same region (like Europe or South Asia) looked very similar to each other. But the map for Europe looked very different from the map for South Asia. This means the "recipe" for getting sick isn't the same everywhere.

2. The Same Ingredient Can Taste Different

The study showed that the same factor can have opposite effects depending on the region.

• The Analogy: Think of "education" like a spice. In one country, adding more education might make the "obesity stew" taste better (increase the risk of obesity). In another country, adding that same spice might make the stew taste worse (decrease the risk).
• The Finding:
  • In South Asia, having more education was actually linked to a higher chance of obesity, diabetes, and high blood pressure.
  • In the Middle East and North Africa, having more education was linked to a lower chance of all those same problems.
  • Age was a big driver everywhere, but in Europe and Central Asia, getting older made the risk of obesity jump up much more dramatically than in other places.

3. The "Hidden" Interactions

The researchers also looked at how two factors work together, like a team.

• The Analogy: Imagine a seesaw. Sometimes, the weight of one person (Age) doesn't matter much unless you know who is sitting on the other side (Sex).
• The Finding: In Europe and Central Asia, the combination of a person's age and their sex was a very strong predictor of obesity. For example, the risk of getting heavier as you age changes differently for men and women in that specific region compared to other parts of the world.

4. What Didn't Change Much

Some things were surprisingly consistent.

• The Analogy: No matter where you go, gravity always pulls you down.
• The Finding: Being older was almost always linked to higher risks of diabetes and high blood pressure, no matter the country. Also, surprisingly, in every single region, people who never smoked or drank were actually more likely to be obese than those who currently did. (The paper notes this is a known puzzle in science, possibly because smoking changes how your body regulates weight).

The Big Takeaway

The main message of this paper is that you can't use a "one-size-fits-all" map to understand heart disease.

If a health official tries to fix a problem in South Asia using the exact same strategy that worked in Europe, it might fail because the "strings" in the web are tied differently. The paper argues that to fix these health problems effectively, we need to look at the specific local web of connections in each region, rather than assuming the world works the same way everywhere.

What the paper does NOT say:

• It does not tell doctors how to treat specific patients.
• It does not claim that education causes obesity in South Asia (it just found a link).
• It does not suggest specific new drugs or diets.
• It simply maps out the differences in how these risk factors are connected across the globe.

Technical Summary

Technical Summary: Global Variation in Cardiometabolic Risk Structures

Problem Statement
Cardiometabolic-based chronic disease (CMBCD) arises from complex, multi-tiered interactions among sociodemographic, behavioral, and metabolic factors. While a consensus set of risk factors is globally recognized, existing epidemiological studies have predominantly focused on individual risk factors or population-attributable fractions (PAF) without adequately modeling the structural interdependencies between these variables across diverse populations. Previous research indicates that the impact of specific risk factors varies by region and income level, yet no study has utilized a network-based approach to analyze how the topology of these risk factor interactions differs globally. This study addresses the gap in understanding how population context influences the structural relationships and interactions driving CMBCD outcomes.

Methodology
The study employed a comparative Bayesian network (BN) analysis using nationally representative cross-sectional data from 48 countries participating in the World Health Organization's "STEPwise approach to noncommunicable disease risk factor surveillance" (STEPS) survey (2009–2024). The total sample size comprised 146,086 participants aged 30 and older.

• Data Harmonization and Variables: Fourteen variables were harmonized across datasets, categorized into a hierarchical structure: non-modifiable sociodemographic (age, sex, urban/rural residence), modifiable sociodemographic (education, household income), behavioral (tobacco, alcohol, diet, physical activity), cardiometabolic drivers (adiposity, diabetes, hypertension, hyperlipidemia), and cardiovascular disease (CVD) history. Variables with >40% missingness were excluded per country; remaining missing data were imputed using a random forest algorithm.
• Network Learning: For each country, a country-specific BN was learned using a hill-climbing algorithm with bootstrap resampling (n=1,000) modified to incorporate survey weights. To ensure causal plausibility, edge blacklists were applied to disallow edges from lower-tier to higher-tier variables and among non-modifiable sociodemographic variables. An averaged consensus network was derived for each country, retaining edges present in >50% of bootstrapped networks.
• Topology Comparison: Structural differences between country networks were quantified using the Structural Hamming Distance (SHD). SHD values were compared across World Bank regions and income levels using unpaired two-sampled Wilcoxon tests.
• Inference and Interaction Analysis: Learned BNs were used for conditional probability queries to assess the relative impact of risk factors on cardiometabolic outcomes. Additionally, interaction effects between risk factor pairs (e.g., age and sex) were evaluated using likelihood ratio tests comparing "full" and "reduced" multinomial regression models fitted to simulated datasets derived from the BNs.
• Sensitivity: Robustness was verified using alternative learning algorithms (tabu search), alternative similarity metrics (product-moment correlation), and alternative imputation methods (MICE).

Key Results

• Regional Structural Similarity: Network structures varied significantly by region (p = 0.023), with networks from the same region being more structurally similar than those from different regions. No significant differences were found based on country income levels (p = 0.91) or the year of data collection.
• Differential Risk Factor Impacts: The influence of sociodemographic and behavioral factors on cardiometabolic outcomes was region-specific.
  • Age: Older age (60+ vs. 30–44) was associated with a markedly higher probability of obesity in Europe and Central Asia (+80%) compared to other regions.
  • Education: Higher education was associated with increased probabilities of obesity, diabetes, and hypertension in South Asia, whereas in the Middle East and North Africa (MENA), it was associated with decreased probabilities of these conditions and hyperlipidemia.
  • Sex and Location: Females showed the largest relative increase in obesity probability across all regions, particularly in Sub-Saharan Africa (+175%). Urban residence was strongly associated with increased obesity in Sub-Saharan Africa and East Asia and Pacific.
  • Behavioral Factors: Never smokers and never drinkers had higher probabilities of obesity compared to current users across all regions. Physical activity showed the strongest protective effect in MENA.
• Interaction Effects: The interaction between age and sex in predicting obesity was significant in the highest proportion of countries within Europe and Central Asia. The interaction between age and sex in predicting hypertension was most prevalent in East Asia and Pacific.

Key Contributions

1. First Global Comparative Network Analysis: This is the first study to learn and compare country-specific Bayesian networks of cardiometabolic risk factors across 48 nations, moving beyond individual risk factor analysis to model complex, multi-tiered interdependencies.
2. Methodological Advancement: The study integrates survey weights into both the structure learning (via weighted bootstrap resampling) and parameter learning stages, addressing a limitation in prior BN studies that often ignored complex sampling designs.
3. Evidence for Regional Tailoring: The findings demonstrate that the structural drivers of CMBCD are not uniform globally. The significant regional variation in network topology and risk factor interactions suggests that generalized prevention models may be insufficient.

Significance and Claims
The paper claims that its results reveal specific regional patterns of multi-tiered cardiometabolic risk structures. It emphasizes that while a consensus set of risk factors drives CMBCD, the inter-relationships and structural pathways through which these factors influence clinical outcomes differ by region. Consequently, the authors argue for the necessity of "regionally tailored public health strategies" rather than the application of generalized, evidence-based models. The study posits that understanding these structural drivers is a critical aspect of "transculturalization"—adapting clinical guidelines to respect cultural and environmental differences. The authors note that while the BNs reflect strong statistical associations rather than definitive causal mechanisms, they provide a robust framework for identifying high-risk groups and locally relevant drivers for intervention.

Limitations Noted by Authors
The authors acknowledge limitations including the cross-sectional nature of the data (precluding longitudinal causal inference), the lack of differentiation between Type 1 and Type 2 diabetes, reliance on self-reported CVD history (potentially less accurate in low-resource settings), and the exclusion of certain variables due to missingness. Furthermore, they note that BN structure-learning algorithms prioritize strong relationships, potentially excluding weaker associations, and that the study uses conditional probability queries to illustrate trends rather than to generate precise metrics like PAF or odds ratios.