A Drug Similarity-Based Bayesian Method for Early Adverse Drug Event Detection.

Abstract

Introduction: Biochemical drug similarity-based methods demonstrate successes in predicting adverse drug events (ADEs) in preclinical settings and enhancing signals of ADEs in real-world data mining. Despite these successes, drug similarity-based ADE detection shall be expanded with false-positive control and evaluated under a time-to-detection setting.

Methods: We tested a drug similarity-based Bayesian method for early ADE detection with false-positive control. Under the tested method, prior distribution of ADE probability of a less frequent drug could be derived from frequent drugs with a high biochemical similarity, and posterior probability of null hypothesis could be used for signal detection and false-positive control. We evaluated the tested and reference methods by mining relatively newer drugs in real-world data (e.g., the US Food and Drug Administration (FDA)'s Adverse Event Reporting System (FAERS) data) and conducting a simulation study.

Results: In FAERS analysis, the times to achieve a same probability of detection for drug-labeled ADEs following initial drug reporting were 5 years and ≥ 7 years for the tested method and reference methods, respectively. Additionally, the tested method compared with reference methods had higher AUC values (0.57-0.79 vs. 0.32-0.71), especially within 3 years following initial drug reporting. In a simulation study, the tested method demonstrated proper false-positive control, and had higher probabilities of detection (0.31-0.60 vs. 0.11-0.41) and AUC values (0.88-0.95 vs. 0.69-0.86) compared with reference methods. Additionally, we identified different types of drug similarities had a comparable performance in high-throughput ADE mining.

Conclusion: The drug similarity-based Bayesian ADE detection method might be able to accelerate ADE detection while controlling the false-positive rate.