Identifying individuals at risk for weight gain using machine learning in electronic medical records from the United States

Aims

Numerous risk factors for the development of obesity have been identified, yet the aetiology is not well understood. Traditional statistical methods for analysing observational data are limited by the volume and characteristics of large datasets. Machine learning (ML) methods can analyse large datasets to extract novel insights on risk factors for obesity. This study predicted adults at risk of a ≥10% increase in index body mass index (BMI) within 12 months using ML and a large electronic medical records (EMR) database.

Materials and Methods

ML algorithms were used with EMR from Optum's de-identified Market Clarity Data, a US database. Models included extreme gradient boosting (XGBoost), random forest, simple logistic regression (no feature selection procedure) and two penalised logistic models (Elastic Net and Least Absolute Shrinkage and Selection Operator [LASSO]). Performance metrics included the area under the curve (AUC) of the receiver operating characteristic curve (used to determine the best-performing model), average precision, Brier score, accuracy, recall, positive predictive value, Youden index, F1 score, negative predictive value and specificity.

Results

The XGBoost model performed best 12 months post-index, with an AUC of 0.75. Lower baseline BMI, having any emergency room visit during the study period, no diabetes mellitus, no lipid disorders and younger age were among the top predictors for ≥10% increase in index BMI.

Conclusion

The current study demonstrates an ML approach applied to EMR to identify those at risk for weight gain over 12 months. Providers may use this risk stratification to prioritise prevention strategies or earlier obesity intervention.