Using Machine Learning to Predict Weight Gain in Adults: an Observational Analysis From the All of Us Research...

Area of Science:

• Public Health
• Biomedical Informatics
• Machine Learning

Background:

• Obesity (body mass index ≥30 kg/m²), a significant public health issue in the U.S., necessitates effective prediction of weight gain risk.
• Current preventative strategies are hindered by the inability to accurately identify individuals most susceptible to weight gain.
• The National Institutes of Health (NIH) All of Us dataset offers a valuable resource for developing predictive models.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting significant weight gain (≥10% total body weight) over a 2-year period.
• To assess whether incorporating behavioral survey data improves the predictive performance of these models compared to using electronic health record data alone.
• To test the hypothesis that combined data sources yield superior weight gain prediction.

Main Methods:

• Utilized the NIH All of Us dataset, including demographics, vital signs, lab results, and survey data (AUDIT, PROMIS scores).
• Developed Elastic Net and XGBoost machine learning models to predict ≥10% total body weight gain within 2 years.
• Employed a 60% training and 40% testing data split, with 10-fold cross-validation for parameter tuning and compared model performance using area under the receiver operating characteristic curves (AUCs).

Main Results:

• The study cohort comprised 34,715 adults aged 18-70 years; 10.4% experienced ≥10% total body weight gain over 2 years.
• XGBoost model achieved an AUC of 0.706, while Elastic Net achieved 0.677, indicating modest predictive performance.
• The inclusion of behavioral survey data did not significantly enhance the predictive accuracy of either machine learning model.

Conclusions:

• Machine learning models demonstrated modest performance in predicting 2-year weight gain, with no significant improvement from incorporating survey data.
• Future research should explore additional variables within the All of Us dataset, such as genomic data, to potentially enhance predictive model accuracy.
• Accurate prediction of weight gain remains a challenge, highlighting the need for further investigation into predictive factors and methodologies.