Flexible statistical approaches for modeling nonlinear relationships in diabetes prediction using splines, Bayesian...

Area of Science:

• Statistical modeling
• Epidemiology
• Machine learning

Background:

• Modeling nonlinear relationships presents a significant challenge in statistical analysis.
• Complex and interacting predictor effects on outcomes require flexible modeling approaches.

Purpose of the Study:

• To compare three flexible statistical methods for capturing nonlinear relationships: restricted cubic spline regression (RCS), Bayesian kernel machine regression (BKMR), and Bayesian additive regression trees (BART).
• To evaluate the performance and clinical interpretability of these methods using the Pima Indians diabetes dataset.

Main Methods:

• Restricted cubic spline regression (RCS) for explicit nonlinear modeling.
• Bayesian kernel machine regression (BKMR) for nonlinear and non-additive effects.
• Bayesian additive regression trees (BART) as a nonparametric ensemble approach.

Main Results:

• RCS and BKMR identified glucose, insulin, age, and skin thickness as significant predictors.
• Bayesian additive regression trees (BART) demonstrated superior predictive performance with an AUC of 97%.
• Predictor-response functions enhanced the clinical interpretability of the models.

Conclusions:

• Flexible methods capable of capturing nonlinear effects can significantly improve prediction accuracy in epidemiological studies.
• Bayesian additive regression trees (BART) offer a powerful approach for complex data analysis in health research.