OBJECTIVES: The Chinese Hong Kong Integrated Modeling and Evaluation (CHIME) is a validated patient-level simulation model derived from a Hong Kong population with prediabetes or type 2 diabetes (T2D) to predict health outcomes. Previous analyses showed that CHIME risk equations possessed satisfactory discrimination but poor calibration for cardiovascular diseases (CVDs) among Taiwanese T2D populations. We therefore sought to recalibrate CHIME risk equations for Taiwanese T2D populations.

METHODS: T2D patients were identified from two leading medical centers in Taiwan, National Cheng Kung University Hospital (NCKUH) and National Taiwan University Hospital (NTUH), during 2013-2019. Study individuals in each hospital were randomly split into training and testing sets to generate the recalibrated risk equations and determine their predictive performance, respectively. Recalibration was implemented by adjusting the overall baseline risk of original risk equations. Further adjustment among low-, middle-, and high-CVD risk patients was performed if the performance after the adjustment of overall baseline risk was not satisfied. Predictive performance was evaluated using calibration ideal slope and Greenwood-Nam-D'Agostino (GND) test (p>0.05) indices.

RESULTS: 32,053 Taiwanese T2D patients were included. Before recalibration, a discrepancy existed between predicted and observed risks in both training sets (slopes: 0.47 and 0.38 for myocardial infarction [MI], 0.72 and 0.80 for stroke, and 0.78 and 0.37 for heart failure [HF] in NCKUH and NTUH, respectively; GND test: p<0.001). After recalibration, the performance of risk equations significantly improved in testing sets of both NCKUH (slopes: 0.98 for MI and 0.94 for HF; GND test: p>0.22) and NTUH (1.10 for MI, 0.87 for stroke, and 1.05 for HF; GND test: p>0.07), except stroke in NCKUH (slope: 1.01; GND test: p<0.01).

CONCLUSIONS: Recalibrated CHIME risk equations are more applicable to Taiwanese T2D patients for predicting CVDs than original risk equations. Other recalibration methods are needed to further improve the stroke risk equation performance.