Application of a machine learning approach to identify predictors of vitamin D status in pregnancy

Abstract Number Theme Presentation Type Cover Approved
0526 Prevalence and risk factors for micronutrient status(deficiency, overload) Poster Not Approved


Abstract Content


In a tropical, urban setting with a high prevalence of vitamin D deficiency, we aimed to examine predictors of vitamin D status among pregnant Bangladeshi women using random forest algorithms.


The Maternal Vitamin D for Infant Growth (MDIG) trial enrolled 1300 pregnant women at 17-24 weeks gestation in Dhaka, Bangladesh. We measured serum 25-hydroxyvitamin D3 (25(OH)D3) by LC-MS/MS for 468 women at enrollment. Risk factors were measured by interview, observation, and a semi-quantitative FFQ and included maternal characteristics, clothing practices, time outside, sunscreen use, season, and diet. Random forest models (ensembles of decision trees) were used to identify the strongest factors among interdependent predictors of vitamin D status; results were compared to conventional multivariable regression.


Women enrolled at a median of 20 weeks gestation and 59% were vitamin D deficient (25(OH)D3 <30 nmol/L). Mean (SD) 25(OH)D was 29.2 (14.9) nmol/L. Almost all women reported spending approximately 1 hour per week outdoors (92%) and having a job under cover/indoors (99%). Median (IQR) intake of vitamin D was 166 (77, 272) IU per week which included 4 (2, 7) servings of fish per week. Using multivariable regression, higher parity and month of assessment were associated with higher vitamin D status. Using random forest, month of assessment and dietary vitamin D intake emerged as the most valuable predictors of higher vitamin D status.


Observed predictors of nutritional status were influenced by the choice of analytical technique. Computational approaches that are distinct from conventional methods may offer new insights in micronutrient research.

Our website uses cookies to ensure you have the best experience.
Please visit our Privacy Policy page for more information.