Class-imbalanced subsampling lasso algorithm for discovering adverse drug reactions.

Area of Science:

• Pharmacovigilance
• Biostatistics
• Data Science

Background:

• Current pharmacovigilance signal detection relies on aggregated data disproportionality analyses.
• Individual spontaneous reports can be analyzed using Bayesian lasso logistic regressions, but face computational challenges and parameter selection issues.
• High-dimensional data in spontaneous reporting present unique challenges for variable selection.

Purpose of the Study:

• To propose an adapted subsampling method for variable selection in high-dimensional pharmacovigilance data.
• To address the imbalance and sparsity inherent in spontaneous adverse event reporting.
• To improve the efficiency and accuracy of safety signal generation.

Main Methods:

• Adapted Stability Selection using an over-sampling scheme for minority classes (adverse events).
• Reduced subsample sizes to manage computational load and data imbalance.
• Compared the proposed method with Stability Selection, gamma-Poisson shrinker, and lasso logistic regression via simulations and empirical study.

Main Results:

• Simulations demonstrated the proposed sampling strategy yields fewer false discoveries and is faster than standard Stability Selection.
• Empirical evaluation on the French pharmacovigilance database showed superior performance in retrieving known safety signals compared to existing methods.
• The method effectively handles large, sparse, and imbalanced spontaneous reporting data.

Conclusions:

• The proposed subsampling method offers a more efficient and accurate approach to pharmacovigilance safety signal detection.
• This method enhances the identification of adverse drug reactions from large-scale spontaneous reporting databases.
• The findings suggest a significant advancement in computational pharmacovigilance methodologies.