Interpretable and Fair Machine Learning for Health-Cost Prediction and Risk Adjustment: A Reproducible Benchmark on Public Data and an Open-Source Toolkit

Abstract

Background. Risk-adjustment formulas set what health plans are paid for each enrollee. A formula that underpays a group gives plans a reason to avoid it, one that responds to recorded diagnoses rewards coding, and one that responds to past use rewards use. Machine-learning methods promise accuracy, but these three properties are rarely evaluated together, on public data, with released code.
Methods. We built a prospective benchmark from five two-year Medical Expenditure Panel Survey panels (43,963 persons, 2018 to 2023) in which year-1 information predicts year-2 total spending. Ten models, including a non-negative least-squares model in the form of a payment formula, gradient boosting and a combined actuarial neural network, were evaluated by cross-validation that keeps primary sampling units whole, with Rao-Wu bootstrap and paired intervals. Outcomes were accuracy, net compensation for nine groups, an accuracy-fairness frontier, SHAP drivers, and the response of payment to simulated coding and to added use.
Results. Gradient boosting reached an out-of-sample R² of 0.187 (95% interval 0.156 to 0.242) against 0.123 for weighted least squares, a paired gain of 0.064 (0.047 to 0.091) that disappeared without prior use and spending. Every model underpaid people who need help with daily activities, by $4,545 to $9,583 per person-year on a mean cost of $31,329. A constrained stacked estimator cut the largest target-group gap to $296 at an R² of 0.171, but both constrained estimators moved $2,242 to $3,362 of overpayment onto enrollees aged 65 and over, and out of fold the constrained gap swung between folds by more than $17,000. A coder targeting each model’s own most lucrative codes raised payment by $11,045 per added code under the payment-form model and $17,469 under the constrained linear formula, against $5,510 under boosting, on the feature set without prior use.
Conclusions. Flexible models are more accurate only when they use prior utilization, which payment formulas exclude. No model pays for functional need unless it observes it. Fairness constraints hold in training but not reliably out of sample, and they move money between groups, so they must be reported for all groups with intervals.