InferBERT

InferBERT infers causal relationships between drug exposure and clinical outcomes from text-based pharmacovigilance data.

Key Features:

• Integration of ALBERT and Do-calculus: Combines A Lite Bidirectional Encoder Representations from Transformers (ALBERT) with Judea Pearl's do-calculus for causal inference on text data.
• Causal Inference Model: Predicts clinical events and identifies their causes from text-based observational data.
• Case Study Evaluations: Evaluated on FDA Adverse Event Reporting System (FAERS) data for analgesics-related acute liver failure and tramadol-related mortalities.
• Performance Metrics: Reported accuracies of 0.78 for analgesics-induced acute liver failure and 0.95 for tramadol-associated deaths.
• Consistency with Clinical Knowledge: Inferred causes were highly consistent with established clinical knowledge.
• Causal Tree Organization: Organizes inferred causal relationships into a causal tree using a recursive do-calculus algorithm.
• Robustness Assessment: Demonstrated high reproducibility and robustness across assessments.

Scientific Applications:

• Pharmacovigilance causal estimation: Establishes causal effects from observational pharmacovigilance datasets such as FAERS.
• Clinical event prediction: Predicts adverse clinical events associated with drug exposures from case report text.
• Drug safety signal evaluation: Supports identification and assessment of potential drug-related adverse event signals for further investigation.
• Understanding intrinsic causality: Enhances interpretation of causal relationships within text-based observational healthcare data.

Methodology:

Integrates transformer-based NLP (ALBERT) with Judea Pearl's do-calculus and applies a recursive do-calculus algorithm to organize inferred causal relationships from FAERS text data.