Drug safety data mining with a tree-based scan statistic.

Area of Science:

• Pharmacovigilance
• Data Mining
• Health Informatics

Background:

• Post-marketing drug safety surveillance aims to detect rare but serious adverse events.
• Traditional methods assess drug-event pairs at a fixed granularity, potentially missing nuanced relationships.
• The optimal granularity for detecting drug-induced events (specific vs. related disorders) is often unknown.

Purpose of the Study:

• To evaluate the tree-based scan statistic as a data mining method for enhancing drug safety surveillance.
• To assess the method's ability to analyze overlapping diagnosis groups at various granularity levels.
• To determine the utility of this approach in identifying potential adverse drug events from electronic health records.

Main Methods:

• Utilized a large electronic health records database (3 million members).
• Applied the tree-based scan statistic to evaluate the safety of antifungal and diabetes drugs.
• Simultaneously assessed overlapping diagnosis groups at different granularity levels, adjusting for multiple testing and covariates (age, sex, health plan).

Main Results:

• Out of 732 evaluated disease groupings, 24 statistically significant signals were identified across 10 non-overlapping disease categories.
• Five signals confirmed known adverse effects.
• Four signals were attributed to confounding by indication, and one signal requires further investigation.

Conclusions:

• The tree-based scan statistic is a viable data mining tool for drug safety surveillance using observational data.
• The identified signals are candidates for further rigorous epidemiological study, not definitive causal links.
• This method aids in generating hypotheses for evaluating drug safety profiles.