Unsupervised Machine Learning Identifies Cardio-Metabolic Obesity Subtypes in Older Chinese

Overview

This study applied unsupervised machine learning to classify cardio-metabolic risk profiles among middle-aged and older Chinese individuals with general and central obesity. Using data from the nationally representative CHARLS cohort, distinct obesity subtypes were identified based on clinically accessible indicators, enabling improved risk stratification beyond BMI alone.

Background

Obesity is a complex chronic disease linked to increased morbidity and mortality worldwide. Traditional classification using BMI does not fully capture the heterogeneity in cardio-metabolic risk associated with obesity. Central obesity, often measured by waist-to-height ratio (WHtR), is particularly relevant in the Chinese population due to its stronger association with non-communicable diseases. Machine learning, especially unsupervised clustering, offers a novel approach to uncover hidden patterns in large clinical datasets to better characterize obesity subtypes and their associated risks.

Data Highlights

The study analyzed data from 7,970 participants aged 45 years or older from the China Health and Retirement Longitudinal Study (CHARLS). Participants had complete anthropometric and laboratory data including BMI, waist circumference, blood pressure, fasting plasma glucose, lipid profiles, and uric acid. The cohort was derived from an initial 17,708 individuals with rigorous data cleaning and exclusion criteria applied to ensure data completeness and quality. Follow-up data from Wave 3 (2015) were also used for model validation.

Key Findings

• Unsupervised machine learning effectively identified distinct cardio-metabolic subgroups within general and central obesity categories in a large Chinese cohort.
• BMI alone was insufficient to differentiate cardio-metabolic risk profiles, highlighting the need for multidimensional assessment including WHtR and biochemical markers.
• Central obesity measured by WHtR showed stronger associations with adverse cardio-metabolic profiles compared to general obesity defined by BMI.
• Data-driven clustering revealed heterogeneity in metabolic health status among obese individuals, suggesting the existence of metabolically healthy and unhealthy obesity phenotypes.
• The identified obesity subtypes can inform precise risk stratification and tailored clinical management strategies in middle-aged and older Chinese populations.

Clinical Implications

Clinicians should consider integrating multiple anthropometric and biochemical indicators beyond BMI to assess obesity-related cardio-metabolic risk more accurately. Employing data-driven subtyping approaches may enhance early identification of high-risk individuals and guide personalized prevention and treatment strategies. This approach supports moving towards precision medicine in obesity management, particularly in populations with high prevalence of central obesity.

Conclusion

Unsupervised machine learning clustering of clinical and biochemical data from a large Chinese cohort revealed distinct cardio-metabolic obesity subtypes, underscoring the limitations of BMI-centric classification. These findings facilitate improved risk stratification and personalized management of obesity in middle-aged and older adults.

References

1. China Health and Retirement Longitudinal Study (CHARLS) -- Dataset and Cohort Description