Associations between Body Mass Index and Lung Function Using Z-Scores: A Nonlinear Relationship and Machine Learning Classification Modeling

Wei Feng¹Fei Lu²Jiangjiang Liu²Yu Zhang²Shiyu Shen²Haitao Ma^2*

• ¹Department of endocrinology, Dushu Lake Hospital Affiliated to Soochow University, Suzhou, China
• ²The Fourth Affiliated Hospital of Soochow University, Suzhou, China

Introduction: This study systematically investigated the relationship between body mass index (BMI) and lung function, incorporating Z-scores, thereby offering a novel approach to lung function management. Methods: Data from the National Health and Nutrition Examination Survey (NHANES, 2007-2012) were utilized, encompassing composite measures of lung function, diet, BMI, smoking history, dust exposure, heart failure, asthma, chronic bronchitis, tuberculosis, a history of thoracic surgery and other relevant covariates. Lung function Z-scores were calculated, and their associations were evaluated using multiple linear regression, logistic regression, and restricted cubic spline models. A total of 12,783 participants were included, with participants categorized into four groups based on forced expiratory volume in one second (FEV1) Z-scores, forced vital capacity (FVC) Z-scores and FEV1/FVC Z-scores: the Z1group, representing the normal lung function group (n = 10,760), the Z2 group, representing the obstructive ventilatory defect group (n = 1,300), the Z3 group, representing the restrictive ventilatory defect group (n = 597), and the Z4 group, representing the mixed ventilatory defect group (n = 126). Subgroup analyses were also performed. We captured the complex relationships between BMI and lung function by developing 22 derived features, employing the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance, and training and comparing seven machine learning algorithms. Results: Among 12,783 participants (mean age 46 years, 51% male), 10,760 had normal lung function, 1,300 had obstructive ventilatory defect (OVD), 597 had restrictive ventilatory defect (RVD), and 126 had mixed defect. BMI demonstrated opposing associations with ventilatory defects: higher BMI was inversely associated with OVD risk (Q4 vs. Q1: OR=0.532, 95% CI 0.418-0.678, P<0.0001), but positively associated with RVD risk (Q4 vs. Q1: OR=2.900, 95% CI 2.708-4.048, P<0.0001). Restricted cubic spline analysis revealed a U-shaped relationship for RVD, with a threshold at 26.39 kg/m². Machine learning models confirmed BMI-related features as the most important predictors, accounting for >32% of total feature importance. Conclusions: This study reveals differential and opposing associations between BMI and ventilatory impairment phenotypes, with higher BMI inversely associated with obstructive defects but positively associated with restrictive defects. Moreover, strong correlations were validated through extensive adjustments and machine learning models.