Drug safety signal detection in a regional healthcare database using the tree-based scan statistic and comparison to 3 other mining methods.

Abstract

To evaluate and compare the relative performance of the tree-based scan statistic (TreeScan) with the crude cohort study, Bayesian confidence propagation neural network (BCPNN) and Gamma Poisson Shrinker (GPS) in detecting statin-related adverse events (AEs) in an electronic healthcare database. Data from a Chinese healthcare database from 2010 to 2016 were evaluated. We identified statin users based on prescription information in their out-/in-patient records, and AEs were defined according to the ICD-10 codes in patients' diagnosis records. TreeScan was applied to detect AE signals related to statin use and was compared with 3 other methods based on sensitivity, specificity, positive predictive value, negative predictive value, accuracy, the Youden index, area under the precision-recall curve and the area under the receiver operating characteristic curve. A total of 224 187 patients were enrolled and divided into 85 758 statin users and 138 429 nonusers. TreeScan generated 29 positive signals, of which 9 were known AEs. The sensitivities of TreeScan, BCPNN and GPS were all 69.2%, which was higher than that of the crude cohort study (46%). The specificity (82.3%), positive predictive value (31.0%), negative predictive value (95.9%), accuracy (81.0%), Youden index (51.5%) and area under the receiver operating characteristic curve (75.8%) of TreeScan were the highest among the 4 methods. TreeScan outperformed the crude cohort, BCPNN and GPS in detecting statin-related AEs in an electronic healthcare database. Therefore, it can be used as a complementary tool for other signal detection methods in drug safety surveillance.