OBJECTIVES: Accurately impute missing service-level medical claim allowed amounts.

METHODS: Line-level allowed amounts were predicted using a Gradient Descent Boosting Regressor machine learning model, trained on a set of independent variables representing provider practice costs, local market payer/provider bargaining dynamics, local market demographics, service type, and billed charge amount. The Gradient Descent Boosting Regressor is a decision tree ensemble algorithm, first developed by Friedman in 1999, that combines the potential for high predictive accuracy with a high degree of flexibility (e.g., it can optimize different cost functions and be extensively tuned). The training data set included 26 million claim-line records, representing insurance segments where reimbursement is subject to rate negotiation: commercial, Medicare Advantage, and managed Medicaid.

RESULTS: The model exhibited strong performance upon validation, with an R² value of .87 and a Mean Absolute Error (MAE) of $13.83. Due to the generalizable, code-agnostic design the model was able to estimate allowed amounts for 99.5% of records in the validation sample.

CONCLUSIONS: The Gradient Descent Boosting Regressor, coupled with a robust and carefully-designed independent and dependent variable table structure, is capable of producing accurate estimates of service-line level medical allowed amounts across procedure codes and three payer types.