Explainable detection of adverse drug reaction with imbalanced data distribution

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Pharmacovigilance

Background:

• Health-related text analysis is crucial for detecting adverse drug reactions (ADRs).
• Data imbalance, where ADR symptoms are minority classes, poses a significant challenge for current models.
• Existing methods like oversampling/down-sampling can disrupt syntactic relationships in sequence labeling.

Purpose of the Study:

• To propose a novel weighted variant of Conditional Random Field (CRF) for data-imbalanced sequence labeling tasks in ADR detection.
• To address the limitations of traditional methods in handling imbalanced datasets for adverse drug reaction identification.
• To evaluate the effectiveness of the proposed weighted CRF model compared to existing sequence labeling approaches.

Main Methods:

• Developed a weighted variant of Conditional Random Field (CRF) to mitigate data distribution imbalances.
• Utilized a CRF model, which captures label relationships between tokens, unlike softmax-based approaches.
• Employed the Locally Interpretable Model-Agnostic Explanations (LIME) algorithm to analyze model performance differences.

Main Results:

• The proposed weighted CRF model effectively alleviates data distribution imbalances between majority and minority classes.
• Experimental results on two distinct ADR tasks demonstrated superior performance compared to previous sequence labeling methods.
• LIME analysis confirmed performance differences, highlighting the benefits of the weighted loss function.

Conclusions:

• The weighted CRF approach offers a robust solution for data-imbalanced sequence labeling in adverse drug reaction detection.
• This method enhances the ability to identify minority class tokens, crucial for accurate ADR identification.
• The proposed model represents a significant advancement in computational pharmacovigilance.