Causality Assessment of Adverse Drug Reaction: Controlling Confounding Induced by Polypharmacy

Area of Science:

• Pharmacovigilance and clinical data analysis
• Computational linguistics and natural language processing
• Medical informatics

Background:

• Post-marketing pharmaceutical surveillance, including pragmatic clinical trials (PCT), is crucial for patient safety.
• Accurately identifying causal drug-adverse drug reaction (ADR) relationships in clinical texts is challenging due to confounding factors from polypharmacy.
• Existing methods struggle to reliably detect drugs causing observed ADRs when multiple medications are involved.

Purpose of the Study:

• To enhance the performance of detecting causal drug-ADR relationships from clinical texts.
• To develop a mechanism for mitigating the impact of confounding on the drug-ADR causality assessment process.

Main Methods:

• A novel analogy-based active voting (AAV) model was proposed to improve the detection of causal drug-ADR pairs.
• The model is inspired by the analogy principle, addressing challenges in polypharmacy scenarios.
• The AAV model aims to improve the ability to detect causal drug-ADR pairs when multiple drugs are prescribed for comorbidities.

Main Results:

• Experimental results demonstrated improved recognition of causal drug-ADR relationships, validated against SIDER.
• The AAV model shows promise in identifying infrequently observed causal drug-ADR pairs, even for less common drugs.
• The study highlights the model's effectiveness in improving the accuracy of causality assessment.

Conclusions:

• The proposed AAV model effectively controls polypharmacy-induced confounding, enhancing ADR causality assessment quality.
• The analogy principle is demonstrated to be applicable and beneficial for assessing drug-ADR causality.
• This research contributes to more reliable post-marketing pharmaceutical surveillance and patient safety.