Adverse drug-event detection using the tree-based scan statistics (TreeScan) and comparison with common mining methods: new user, propensity score-matched cohorts.

purposeTree-based scan statistics (TreeScan) combined with a new user propensity score (PS)-matched cohort design (PS-TreeScan) enhances post-marketing drug safety surveillance by adjusting for confounding biases. However, this method has not been compared with other signal mining techniques or used in post-marketing drug safety studies in China. We evaluate the effectiveness of the PS-TreeScan method in identifying statin-associated adverse events (AEs) by comparing it with three established signal detection methods using a Chinese real-world database.

methodsWe used data from the Yinzhou District Medical Database, encompassing patients with hypertension (2010–2016). Statin exposure and AEs were determined via outpatient/inpatient prescriptions and using ICD-10 codes in similar settings, respectively. A new user design with a 1:1 PS matching was implemented. Standard positive and negative signals were obtained from published systematic reviews, meta-analyses, and summary of product characteristics (SPC). PS-TreeScan was compared with three mining methods—incident rate ratio (crude cohort), Bayesian Confidence Propagation Neural Network (BCPNN), and Gamma Poisson Shrinker (GPS)—to identify statin-related AEs. Evaluation indices were calculated using the diagnostic test evaluation method, and area under the receiver operating characteristic curve (AUC) values were compared to assess differences.

resultsPS-TreeScan identified 15 positive signals (P < 0.05), including 8 true positives. PS-TreeScan’s sensitivity was equivalent to those of BCPNN and GPS (62%) but higher than that of the crude cohort method. TreeScan AUC values were significantly higher at 77.7% (95% confidence interval: 63.7%–91.6%).

conclusionCompared to the original data, matched data effectively reduced false positives and improved the AUC for all methods. The PS-TreeScan method outperformed the traditional methods, thus it is able to supplement other mining methods for active adverse drug reaction monitoring.