Claims-based Model to Predict Obesity in Older Patients

Background: In epidemiological studies, obesity is often an important confounder of the studied associations, and thus a measurement of obesity is required for confounder adjustment. However, obesity weight status, as measured by body mass index (BMI), is often unavailable or under-recorded in administrative claims data.

Objectives: To develop and validate a claims-based algorithm that classifies patients into obesity categories defined by BMI.

Methods: Using Medicare claims data linked to two multi-center electronic health records (EHR) systems in Massachusetts, United States, we identified a cohort of patients with an available BMI measurement (measurement date = cohort entry date [CED]) in the EHR dataset. Patients were required to have Medicare enrollment 90 days prior and 90 days after the CED. Patients were excluded if they had missing information on age or gender, or a previous bariatric surgery procedure, limb amputation or any cancer. A prediction model was constructed using the EHR system 1 (training set) and validated using EHR system 2 (validation set). We used least absolute shrinkage and selection operator (LASSO) regression to select predictors from 137 prespecified variables. The predicted BMI based on a generalized linear model with LASSO-selected variables was used to classify patients with BMI ≥ 25, BMI ≥ 30 and BMI ≥ 40.

Results: The study cohort contained 145,948 patients (mean age 74; female 61 %). The final model comprised 97 variables. The main predictors of BMI were BMI-related diagnosis codes, as well as age, sex, race, cardiovascular and antidiabetic drugs, and several obesity-related comorbidities (e.g., Type 2 Diabetes, cancer, hypertension). In the validation set, the final model correctly identified 65.6% (95% confidence interval [CI]: 65.2% to 66.0%) of patients with BMI ≥ 25, 75.3% (95% CI: 74.9%-75.7%) with BMI ≥ 30 and 95.9% (95% CI: 95.8% to 96.1%) with BMI ≥ 40. The positive predictive value of the model was 81.5% (95% CI: 81.0% to 81.9%), 80.6% (95% CI: 79.6% to 81.6%) and 64.7% (95% CI:48.6% to 81.6%) for BMI ≥ 25, BMI ≥ 30 and BMI ≥ 40, respectively.

Conclusions: The prediction model proposed in this study can be used to identify patients with obesity classes in terms of BMI categories in claims databases among patients 65 years of age or older when BMI measurements are not available. The claims-based BMI phenotypes can be used for confounding adjustment or defining subgroups in pharmacoepidemiology studies.