2.5. Establish the trajectory model

To categorize the trend of BP over time, we applied GBTM. The longitudinal SBP and DBP data were fitted as trajectories in a multivariate censored normal (CNORM) model (29). Multi-trajectories of SBP and DBP mean that each individual in a group has two trajectory curves: one for SBP and the other for DBP. Starting with a one-group model, we evaluated the optimal number of groups and polynomial type of each group trajectory (intercept, linear, or quadratic) with the following criteria: (1) Bayesian information criterion (BIC), where a value closer to 0 indicates a better fit model; (2) the value of the Bayesian factor logarithm, which is approximately two times the difference in BIC between the two compared models; (3) allocation of at least >5% of the total patients in each trajectory group in a model; and (4) higher average probability of final group membership across the trajectory groups.

As the number of trajectory groups was >4, although the Bayesian information criterion was closer to 0, the value of Bayesian factor logarithm was bigger (Supplementary Table S2), and the number of participants in one group was <5% of total. Finally, four distinct trajectories with regard to changes in both SBP and DBP were identified: 967 (7.16%), 7,450 (55.17%), 4,356 (32.26%), and 731 (5.14%) patients were assigned to the class 1, class 2, class 3, and class 4 trajectory groups, respectively. Supplementary Tables S2 and S3 present related parameters of the optimal multivariate trajectories of SBP and DBP after several attempts.