Improved diagnosis-medication association mining to reduce pseudo-associations

Area of Science:

• Medical Informatics
• Pharmacovigilance
• Data Mining

Background:

• Association rule mining in medicine often yields spurious disease-medication (DM) associations.
• Existing methods struggle to differentiate true DM relationships from false positives.

Purpose of the Study:

• To develop and evaluate a modified association rule mining algorithm for identifying accurate real-world DM associations.
• To assess the algorithm's effectiveness in reducing pseudo-associations.

Main Methods:

• Utilized outpatient claims data (2011-2015) from Taiwan's Health and Welfare Data Science Center.
• Employed a modified lift (Q-value) metric (Q2) to quantify DM associations, comparing it to the original (Q1).
• Validated 1000 DM pairs flagged as potential pseudo-associations by pharmacists.

Main Results:

• Identified over 3.1 million unique DM pairs, with significant numbers flagged as potential pseudo-associations by the original algorithm (Q1+).
• The modified algorithm (Q2) correctly identified 93.7% of these as pseudo-associations.
• Specific drug classes (M, H, B) showed the highest rates of plausible associations missed by the modified algorithm.

Conclusions:

• The modified algorithm accurately detects pseudo-associations missed by standard methods.
• This approach can enhance secondary databases for real-world prescribing pattern research.
• Improved identification of DM associations can contribute to enhanced drug safety.