Divergent Cardiometabolic Comorbidity Patterns Between Physically Active and Inactive US Adults With Prediabetes: An Association Rule Analysis and Explainable Machine Learning Study Using NHANES 2011–2018

Technical Summary: Divergent Cardiometabolic Comorbidity Patterns Between Physically Active and Inactive US Adults With Prediabetes

Problem Statement
Prediabetes affects approximately 38% of US adults and frequently co-occurs with a constellation of cardiometabolic conditions, including hypertension, obesity, dyslipidemia, and liver enzyme abnormalities. While physical activity (PA) is a cornerstone of diabetes prevention, existing literature has largely focused on the reduction of individual disease prevalence rather than how PA status influences the architecture of multimorbidity patterns. Traditional regression-based epidemiological methods often assume independence among predictor variables and fail to capture the complex, non-linear co-occurrence of multiple conditions. Furthermore, no study had systematically applied data-driven approaches to examine PA-stratified comorbidity patterns specifically in US adults with prediabetes, nor had any study combined association rule mining with explainable machine learning to provide complementary analytical perspectives on this relationship.

Methodology
This study utilized a large, nationally representative dataset from the National Health and Nutrition Examination Survey (NHANES) 2011–2018, comprising 10,891 US adults aged ≥20 years. The analytic sample included 3,228 individuals meeting American Diabetes Association (ADA) criteria for prediabetes and 1,154 with Type 2 Diabetes (T2D).

The study employed a dual-method analytical framework:

1. Association Rule Mining (ARM): Using the Apriori algorithm, the researchers mined frequent co-occurrence patterns among five cardiometabolic conditions (obesity, hypertension, dyslipidemia, liver enzyme abnormality, and T2D). The analysis was stratified by PA status, defined according to WHO guidelines (≥600 MET-minutes/week as "active" vs. <600 as "inactive"). Rules were evaluated using support, confidence, and lift metrics, with minimum thresholds set at support ≥3%, confidence ≥20%, and lift >1.2.
2. Explainable Machine Learning (XGBoost-SHAP): An XGBoost classifier was trained to predict T2D status. The SHapley Additive exPlanations (SHAP) framework was applied to decompose predictions and quantify the relative importance of features (including PA status, glycemic status, and specific comorbidities) in a stratified manner.

The study also conducted a cross-population validation by replicating the analytical framework in the T2D subgroup and comparing findings with a recent Korean KNHANES study (Huang et al., 2026) to assess consistency across different glycemic stages and populations.

Key Results

• Divergent Rule Counts: Physically active adults with prediabetes exhibited significantly more diverse comorbidity patterns than inactive counterparts. The active group generated 30 association rules, compared to only 12 rules in the inactive group (a 2.5-fold difference). This pattern was robustly validated in the T2D subgroup, where active individuals showed 54 rules versus 24 in inactive individuals.
• Compositional Differences: The composition of the rules differed markedly. The active group displayed diverse clustering involving dyslipidemia, hypertension, and liver enzyme abnormalities. In contrast, the inactive group's patterns were highly homogeneous and dominated by an obesity-centered axis (e.g., obesity-dyslipidemia clustering).
• Strength of Clustering: The mean lift was higher in the active group (1.43 vs. 1.27 in prediabetes), indicating stronger co-occurrence clustering among the identified disease combinations in the active population.
• Feature Importance (SHAP): In the XGBoost model predicting T2D, prediabetes status was the dominant predictor (SHAP = 2.678), followed by age and obesity. PA status itself ranked 7th out of 8 features (SHAP = 0.110), suggesting that while PA engagement contributes to risk prediction, its individual predictive power is modest compared to direct metabolic markers.
• Cross-Population Discrepancy: The study found a directionally opposite pattern compared to the Korean KNHANES study (which reported fewer rules in active groups). The authors attribute this to differences in baseline metabolic burden (higher BMI in US vs. Korean populations), healthcare access/diagnostic surveillance, and cultural reporting biases.

Key Contributions

1. Novel Methodological Integration: This is the first study to combine association rule mining with explainable machine learning (XGBoost-SHAP) to investigate PA-stratified comorbidity patterns. This approach allows for the discovery of hidden multivariate patterns (via ARM) while simultaneously quantifying the marginal contribution of individual features (via SHAP).
2. Systems-Level Insight: The study challenges the conventional view that PA simply reduces disease prevalence. Instead, it demonstrates that PA engagement is associated with a fundamentally different, more heterogeneous comorbidity architecture. The authors argue that PA acts as a "systemic modifier" of disease interdependence rather than merely a direct predictor of disease status.
3. Population-Specific Evidence: By highlighting the discrepancy between US and Korean findings, the study underscores the critical importance of context-specific comorbidity research, cautioning against the uncritical generalization of lifestyle-comorbidity relationships across populations with different body composition profiles and healthcare systems.
4. Validation Across Glycemic Stages: The findings were consistently validated across both prediabetes and T2D populations, confirming the robustness of the observed patterns.

Significance and Claims
The paper claims that physically active US adults with prediabetes exhibit more numerous and diverse cardiometabolic comorbidity patterns than inactive individuals. This finding suggests that PA engagement reshapes the multimorbidity landscape at a systems level, creating a more complex but potentially distinct architecture of disease co-occurrence.

The authors posit that this has significant implications for clinical practice and public health:

• Screening Strategies: PA status could inform targeted screening; inactive individuals may require focused surveillance for obesity-centered clusters, while active individuals may need broader surveillance across a more heterogeneous disease landscape.
• Intervention Framing: Public health strategies should frame PA not just as a tool to reduce individual risk factors, but as a systemic intervention that restructures how conditions co-occur, potentially reducing overall disease management complexity.

The study concludes that while the specific direction of the PA-comorbidity relationship may vary by population (as seen in the US vs. Korea comparison), the use of advanced data mining techniques is essential for uncovering these complex, non-linear relationships that traditional epidemiological methods miss.