Applying machine learning approaches for predicting obesity risk using US health administrative claims database

Area of Science:

• Health Informatics
• Data Science in Healthcare
• Public Health Surveillance

Background:

• Body mass index (BMI) recording is often incomplete in US administrative claims databases.
• Validating BMI-related diagnosis codes and predicting obesity status are crucial for accurate health burden assessment.

Purpose of the Study:

• To validate the sensitivity and positive predictive value (PPV) of BMI-related diagnosis codes.
• To apply machine learning (ML) models to predict obesity status using US claims data.

Main Methods:

• Retrospective analysis of 692,119 individuals from January 2013 to December 2019 using MarketScan Explorys Claims-EMR data.
• Comparison of claims-based obesity status with EMR-based BMI (gold standard) to assess code accuracy.
• Training of logistic regression, penalized logistic regression, extreme gradient boosting (XGBoost), and random forest models using insurance claims features.

Main Results:

• Claims data showed high PPV (85.4-89.2%) but low sensitivity (16.8-44.8%) for obesity diagnosis codes.
• XGBoost demonstrated superior performance in predicting obesity, achieving the highest area under the curve (AUC) of 79.4%.
• The number of obesity diagnoses and inpatient obesity diagnoses were key predictors; XGBoost achieved an AUC of 74.1% without explicit obesity codes.

Conclusions:

• Obesity prevalence is under-reported in administrative claims databases.
• Machine learning models show significant promise for enhancing obesity prediction accuracy, even without explicit obesity diagnosis codes.
• Improved obesity prediction can aid practitioners and payors in estimating the obesity burden and identifying treatment needs.