Oral Presentation Australian and New Zealand Obesity Society Annual Scientific Conference 2023

Applying Machine Learning Methodology for Early Prediction of Adolescent and Adult BMI (98112)

Rae-Chi Huang 1 , Fuling Chen 2 , Lawrence J Beilin 3 , Trevor A Mori 3 , Phillip Melton 4
  1. Nutrition and Health Innovation Research Institute, Edith Cowan University, Perth, WA, Australia
  2. Telethon Kids Institute, University of Western Australia, Perth, WA, Australia
  3. Medical School, University of Western Australia, Perth, WA, Australia
  4. University of Tasmania, Hobart, Tasmania, Australia

Background:  

Obesity is costly with antecedents in the perinatal and preschool years.  Being able to predict risk for adolescent/adult obesity before children enter school and stratifying early prevention is a worthwhile public health pursuit. 

Aim:

To predict childhood, adolescent and adult Body Mass Index (BMI) (ages 8 to 26 years) from data collected from the mothers in the prepregnancy and pregnancy periods and from the offspring in preschool years (0-5 years old). 

Methods:

Raine Study recruited 2900 pregnant women, of whom 2868 offspring were followed up between 8-26 years old.  191 predictors from pre-pregnancy, pregnancy and preschool years were utilised.  Exclusions included major congenital anomalies or multiparity.  The primary outcomes were offspring BMI at 8, 10, 13, 16, 20, 23 and 26 years-old.

Preprocessing and feature selection were applied.  Data was split into training and testing sets. Machine learning models (Extreme Random Forest (ERF), Elastic Net (EN), Lasso and Ridge) were applied.  To investigate the important factors that the models used to predict the outcome, ERF modelling provided weightings. Python package scikit-learn version 1.0.1 was used. 

Results:

The prediction of childhood BMI (8 and 10 years old) (R2=0.6-0.8) is more precise than the prediction of adolescent BMI (13 and 16 years) (R2=0.4-0.5), which in turn is more precise than adult BMI (20, 23, and 26 years) (R2=0.2-0.3). The best model predicted BMI at 8 years using ERF with R2 =0.765.    When predicting BMI at all ages, anthropometry at 5 years old (38-85% weighting) and maternal weight dominated, carrying the greatest weighting.

Conclusion:

Modest prediction of adolescent BMI can be made from preschool and perinatal factors with diminishing prediction of adult BMI.  The dominant predictors universally across timepoints persisting to adulthood, are anthropometry at 5 years-old and maternal weight, thus supporting the imperative to address preschool obesity trajectories <5 years-old.