DeepCausality: A general AI-powered causal inference framework for free text: A case study of LiverTox

Area of Science:

• Computational biology
• Artificial intelligence
• Causal inference

Background:

• Causality is crucial across sciences, but manual text analysis is inefficient.
• Automated methods are needed for large-scale causal factor identification from free text.

Purpose of the Study:

• To introduce DeepCausality, a general framework for empirical causal inference from free text.
• To demonstrate DeepCausality's application in identifying drug-induced liver injury (DILI) causal factors.

Main Methods:

• Integrated AI language models, named entity recognition, and Judea Pearl's Do-calculus.
• Applied the framework to the LiverTox database for DILI causal term extraction.
• Generated a causal tree for patient stratification and severity scoring.

Main Results:

• Achieved 0.92 accuracy and 0.84 F-score for DILI prediction.
• 90% of identified causal terms aligned with American College of Gastroenterology guidelines.
• High concordance (0.91) between DeepCausality and expert iDILI severity scores.

Conclusions:

• DeepCausality offers a promising automated solution for causality assessment in free text.
• The framework aids in identifying clinical causal terms and assessing disease severity.
• Publicly available code facilitates broader application in scientific research.