Presentation (CTI)

OBJECTIVES: This study aimed to develop and interpret risk adjustment models for outpatient care capitation payments based on family doctor system in China using machine learning algorithms.
METHODS: This study utilized medical insurance claims data and family doctor registration database from a provincial capital city in eastern China between January 1, 2018, and December 31, 2019. Participants were individuals continuously registered with a same primary medical institution during 2018~2019, and utilized outpatient visits in 2018. We developed prospective risk adjustment models using risk adjusters in 2018 to predict total outpatient costs in 2019. K-means clustering and Lasso regression were applied to identify important previous health conditions. Linear regression, random forest, eXtreme Gradient Boosting and Light Gradient Boosting Machine (LightGBM) were used to predict outpatient costs. Shapley additive explanations were employed to interpret model outputs. Under- and overcompensation across costs groups were assessed to reflect potential risk selection incentives.
RESULTS: A total of 950156 individuals were included, with a mean age of 49.15 years. The models incorporated 59 risk adjusters, including age, sex and their interactions, type of basic medical insurance, Charlson Comorbidity Index (CCI), and 54 diagnosed health condition clusters. All three machine learning models outperformed linear regression, with LightGBM achieving the best performance across all metrics. The most important predictors were type of medical insurance, CCI, previous health conditions such as circulatory diseases (e.g. hypertension), cancers, metabolic diseases (e.g. diabetes). Compared with linear regression, machine learning models reduced under-compensation among high-cost patients but increased over-compensation among low-cost patients.
CONCLUSIONS: Machine learning provides improved prediction of outpatient costs compared with linear regression. These findings offer methodological guidance for the development and interpretation of risk adjustment models in China and highlight the need for further investigation into the models’ incentives for risk selection.