A data-driven method to detect adverse drug events from prescription data

Area of Science:

• Pharmacovigilance and Pharmacoepidemiology
• Data Mining and Machine Learning in Healthcare
• Computational Drug Safety

Background:

• Clinical trials often lack sufficient sample sizes to detect rare Adverse Drug Events (ADEs).
• Post-marketing surveillance is crucial for identifying drug safety issues in diverse patient populations.
• Existing methods for ADE detection often rely on Electronic Health Records (EHRs) with diagnostic information.

Purpose of the Study:

• To develop and evaluate a data-driven method for detecting potential ADEs from pure prescription data.
• To identify ADE-associated prescription sequences indicating a potential causal link between a drug and an adverse event.
• To assess the efficacy of the proposed method in identifying both known and novel ADEs.

Main Methods:

• Utilized constrained sequential pattern mining to discover prescription sequences indicative of ADEs.
• Integrated domain constraints to filter out non-relevant signals and reduce noise.
• Applied an adapted Self-Controlled Case Series model to rigorously evaluate potential ADE signals.
• Leveraged a large-scale Australian prescription dataset (over 50 million records, 2 million patients).

Main Results:

• The proposed method successfully identified known ADEs and flagged suspicious, potentially novel ADEs.
• Achieved a high detection rate of positive adverse events (67.4%) compared to a gold standard.
• Demonstrated a low false positive rate (8.78%), indicating high precision.
• Effectively utilized pure prescription data without requiring additional clinical information like diagnoses or symptoms.

Conclusions:

• The data-driven approach using prescription sequences is effective for post-marketing ADE detection.
• This method offers a valuable alternative or supplement to traditional pharmacovigilance systems.
• The findings highlight the potential of mining prescription data for enhancing public drug safety.