3.1 Study population and characteristics
After applying the eligibility criteria, the derivation cohort included 8,176 cases of serious GI complications from a total of 2,004,031 patients who used NSAIDs for more than 30 consecutive days and 81,760 matched controls (Fig 1). The prevalence and rate of serious GI complications were 0.41% and 4.85 per 1000 person-years (95% CI, 4.75–4.96), respectively, in the total derivation cohort. The mean age was 75.9 years, and 71.6% were female; 1.3% of the development cohort had a history of GI bleeding or perforation and 27.6% had a history of GI ulcer without bleeding or perforation (Table 1).
3.2 Model development and performance
The predictors included in the multivariate analysis are presented in Table S2. Predictive risk factors included in the final model after parameter selection were as follows: age, sex, high-dose NSAIDs, type of NSAIDs, GPA with NSAIDs, concomitant antithrombotic agents, glucocorticoids, SSRI, history of GI bleeding or perforation, severe renal disease, and liver cirrhosis. GPA concomitant with NSAIDs included only PPI and H2RA. Misoprostol use did not reach statistical significance due to its infrequent use. The scores and adjusted OR for each risk factor are shown in Table 2. The points from -5 to 18 were assigned for each risk factor. For the prediction of serious GI complications, the cutoff score for classifying low-risk vs. high-risk patients was 16 (Youden Index, 0.42; sensitivity, 0.66; specificity, 0.76; Table 4).
Results from the Hosmer–Lemeshow test for the total score calculated using the risk scores showed an AUROC of 0.78 (95% CI, 0.77–0.79) from the internal training dataset and 0.77 (95% CI, 0.76–0.78) from the internal hold-out validation dataset, which considered acceptable discrimination (Table 3) [17]. The calibration curves of the internal hold-out validation dataset demonstrated highly accurate calibration (calibration slope β =1.11), and the risk of GI complications increased as the prediction scores increased (Fig S1).