Associations of dietary patterns with serum 25(OH) vitamin D and serum anemia related biomarkers among expectant mothers: A machine learning based approach

Background: Machine learning algorithms (MLA) gained prominence in nutritional epidemiology for analyzing dietary associations and uncovering intricate patterns within data. We explored dietary patterns associated with serum iron biomarkers and vitamin D among pregnant women, utilizing MLA to perform predictive analyses. Methods: The cross-sectional study utilized a secondary dataset from the Nationwide Nutrition and Health Survey in Taiwan, and 1,423 expectant mothers were recruited. Dietary patterns were predicted using K-means cluster analysis on semiquantitative food frequency data. Associations between serum biomarkers and dietary patterns were analyzed using binomial logistic regression, adjusting for sociodemographic and dietary variables. MLA including support vector machine, K-nearest neighbor, naive Bayes, random forest, and decision tree were applied to predict the accuracy of the dietary patterns in improving anemia-related biomarkers. Results: The K-means clustering identified two dietary patterns: LP + LA (low plant, low animal) and MP + LA (moderate plant, low animal). Logistic regression revealed that expectant mothers following the MP + LA pattern had a lower likelihood of low serum iron (OR = 0.45, 95 % CI 0.34–0.60) and ferritin (OR = 0.27, 95 % CI 0.21–0.36), but a higher likelihood of low 25(OH) vitamin D. MLA models demonstrated 70 %–76 % accuracy in identifying dietary pattern associated with improvement in serum iron and ferritin levels. Conclusions: The MP + LA dietary pattern exhibits a positive association with serum iron biomarkers and a negative association with 25(OH) vitamin D. Machine learning models demonstrate comparable predictive accuracy, highlighting their utility in nutritional epidemiology for identifying dietary patterns and their relationships with biochemical markers.