InferBERT: A Transformer-Based Causal Inference Framework for Enhancing Pharmacovigilance.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Causal Inference

Background:

• Transformer-based language models excel in NLP but lack causal inference capabilities.
• Causality is crucial for pharmacovigilance and healthcare decision-making.
• Current models cannot determine the cause of clinical outcomes.

Purpose of the Study:

• To develop an innovative causal inference model, InferBERT.
• To integrate ALBERT and Judea Pearl's Do-calculus for pharmacovigilance.
• To address the need for causality inference in clinical outcome analysis.

Main Methods:

• Proposed InferBERT model integrating ALBERT and Do-calculus.
• Evaluation using FDA Adverse Event Reporting System (FAERS) data.
• Case studies: Analgesics-related acute liver failure and Tramadol-related mortalities.
• Causal tree generation using a recursive do-calculus algorithm.
• Robustness assessment for reproducibility.

Main Results:

• InferBERT achieved accuracies of 0.78 for acute liver failure and 0.95 for mortality.
• Inferred causes aligned with existing clinical knowledge.
• Causal trees enhanced understanding of causality.
• High reproducibility demonstrated through robustness assessment.

Conclusions:

• InferBERT successfully predicts clinical events and infers their causes.
• The model establishes causal effects from text-based observational data.
• InferBERT is a promising tool for understanding intrinsic causality in healthcare.