Identifying payment-driven patient selection using machine-learning-predicted cost risk: evidence from neonatal care in China

The global pursuit of Universal Health Coverage (UHC) has spurred a worldwide shift towards case-based payment systems to control costs and expand health coverage. However, payment reform without adequate risk adjustment may incentivize hospitals to avoid high-cost-risk patients, potentially threatening health equity. China's innovative Diagnosis-Intervention Packet (DIP) reform represents a major, globally-relevant experiment in implementing a simplified case-based system at scale. This study evaluates its impact on patient selection for a vulnerable pediatric population.

The study included 157,739 neonatal jaundice hospitalizations, recorded between January 2018 and September 2023. Six candidate cost-prediction models were evaluated using pre-reform data, and the selected model was used to predict patients' cost risk. Patients with predicted costs in the top quartile were classified as high-cost-risk patients, providing a measurable indicator of patient selection. We subsequently applied a difference-in-differences(DID) analysis to estimate changes in the proportion of high-cost-risk admissions following DIP implementation.

Among 6 machine learning models, eXtreme Gradient Boosting (XGBoost) achieved superior accuracy in predicting inherent cost risk (RMSE = 314.14 ± 13.28, bias=-0.01 ± 4.56). DIP implementation was associated with a significant reduction in the proportion of high-cost-risk neonates by 8.9% points (DID = -0.089; 95% CI: -0.126 to -0.053; P < 0.001). Subgroup analyses revealed the strongest effects in hospitals with medium competition strength (DID = -0.116; 95% CI: -0.186 to -0.046; P < 0.001), tertiary hospitals (DID = -0.105; 95% CI: -0.143 to -0.068; P < 0.001) and public hospitals (DID = -0.094; 95% CI: -0.132 to -0.056; P < 0.001).

China's experience provides a critical cautionary tale for the global health community. Case-based payment reform without adequate risk adjustment creates significant and measurable health inequity by systematically reducing access to healthcare for the most clinically vulnerable patients, threatening the foundational goal of equitable access under UHC. The machine-learning-predicted cost-risk measure provides a scalable tool for identifying and monitoring payment-driven patient selection across different policy settings.