Association of TyG-RFM with depressive symptoms: a nationally representative study enhanced by interpretable machine learning.

BACKGROUND: The association between the triglyceride-glucose index-relative fat mass (TyG-RFM) and prevalent depressive symptoms remains unclear. This study aimed to examine this association and to develop an interpretable machine-learning model for screening-oriented assessment of depressive symptoms in US adults. METHODS: A total of 12,600 participants from the National Health and Nutrition Examination Survey 2005-2018 were included in this study. Weighted logistic regression, restricted cubic spline (RCS), and subgroup were performed to evaluate the association between TyG-RFM and depressive symptoms. In addition, nine machine-learning models were developed and internally evaluated to estimate the individualized probability of depressive symptoms. RESULTS: Weighted multivariable regression analysis showed that TyG-RFM was positively associated with depressive symptoms (OR = 1.97, 95% CI: 1.10-3.53), and RCS analysis showed a similar positive relationship. In the machine-learning analysis, LightGBM showed the best overall internal performance based on five-fold cross-validated out-of-fold estimates, with an AUC of 0.747 (95% CI: 0.732-0.762), the lowest Brier score (0.070), and the greatest net benefit on decision curve analysis. SHAP analysis showed that TyG-RFM was the most influential feature in the LightGBM model for estimating the probability of depressive symptoms. CONCLUSIONS: TyG-RFM was positively associated with depressive symptoms and was identified as the most important feature in the best-performing LightGBM model, with the accompanying online calculator providing an exploratory approach for the assessment of depressive symptoms. However, longitudinal studies and external validation are still needed before broader clinical application.