Supplementing claims data analysis using self-reported data to develop a probabilistic phenotype model for current...

Area of Science:

• Health Informatics
• Epidemiology
• Data Science

Background:

• Smoking status is often missing in US health insurance claims data.
• Accurate smoking data is crucial for epidemiological studies and confounder adjustment.
• The IBM MarketScan Commercial database offers a potential source for smoking status imputation.

Purpose of the Study:

• To develop a generalizable smoking status phenotype model using US claims data.
• To investigate the utility of a subset of patients with self-reported smoking status for model training.
• To create a model that calculates the probability of being a current smoker.

Main Methods:

• A subset of 1,966,174 patients with linked health risk assessments was used.
• A regularized logistic regression model, Current Risk of Smoking Status (CROSS), was trained.
• CROSS utilized 53,027 covariates from the prior 365 days, including demographics, conditions, drugs, and procedures.

Main Results:

• The CROSS model achieved an internal AUC of 0.76 and was well-calibrated.
• External validation across three US claims databases yielded AUCs between 0.82 and 0.87.
• The model demonstrated transportability across different claims data sources.

Conclusions:

• The CROSS model effectively predicts current smoking status from prior year claims data.
• CROSS can be implemented with OMOP common data model-mapped US insurance claims.
• This model is valuable for imputing smoking status in epidemiological research where it's a known confounder.