The relationship between number of pregnancies and serum 25-hydroxyvitamin D levels in women with a prior pregnancy: a cross - sectional analysis, machine learning - based prediction model, and SHAP - assisted feature importance evaluation.

Background: The primary aim of this study is to explore the association between gravidity and serum 25-hydroxyvitamin D [25(OH)D] levels in women, as existing research rarely addresses gravidity's cumulative impact on maternal vitamin D status. Secondarily, it seeks to develop and evaluate a machine learning model for predicting vitamin D insufficiency (serum 25(OH)D < 50 nmol/L) using reproductive data (including gravidity) and biochemical indicators, with contribution analysis in the model further validating this relationship, thereby translating the findings into a clinically useful tool. Methods: The study included 8,003 parous women from the NHANES survey conducted between 2011 and 2018, excluding those with missing data on vitamin D or gravidity. For the primary objective, we employed covariate-adjusted linear regression analyses to examine the relationship between gravidity and serum 25(OH)D levels. Three hierarchical models were constructed: Model 1 (unadjusted); Model 2, adjusted for age and race/ethnicity; and Model 3, adjusted for all potential confounders (including body mass index, blood urea nitrogen, glycated hemoglobin, and diabetes status). For the secondary objective of model development, multiple regression analysis and six machine learning algorithms (including XGBoost and Random Forest) were employed. These algorithms are well-suited to handle mixed-type biomedical data (e.g., continuous biochemical indices and categorical reproductive factors), aligning with the characteristics of the dataset in this study. The dataset was subsequently split into a training set and a validation set at a 70:30 ratio. Results: The study found that each additional pregnancy was associated with a 0.6 nmol/L decrease in 25(OH)D concentration (P<0.001). For the secondary objective of predictive modeling, the XGBoost algorithm showed better performance in clinically predicting vitamin D levels, with an area under the receiver operating characteristic curve (AUC) of 0.73, which was superior to multiple regression analysis and the other five machine learning algorithms (including Random Forest, Logistic Regression, Support Vector Machine, Decision Tree, and Naive Bayes) and demonstrated greater efficacy in identifying low serum vitamin D levels. Key features contributing significantly to the model included age, body mass index (BMI), and blood urea nitrogen. Conclusion: In women with a prior pregnancy, an independent inverse association was observed between gravidity and vitamin D status and the XGBoost algorithm demonstrated superior performance in clinically predicting vitamin D levels using common blood test results which may facilitate timely detection and intervention for low serum vitamin D.