CauRuler: Causal irredundant association rule miner for complex patient trajectory modelling

Area of Science:

• Causal inference in healthcare
• Machine learning for clinical data analysis
• Association rule mining

Background:

• Identifying causal associations is key for clinical trials and understanding health status.
• Existing machine learning methods struggle with confounder control and generalizable causal rules in healthcare.
• Prior knowledge of causal paths can help prevent adverse health conditions.

Purpose of the Study:

• To present and evaluate CauRuler, a novel approach for causality discovery from association rules.
• To develop an algorithm suitable for large-scale health databases with robust confounder control.
• To generalize causality using anti-monotone properties for complex causal path identification.

Main Methods:

• CauRuler extends association rule mining with an irredundancy property for rule set reduction and generalization.
• A pruning strategy is employed to reduce the association rule set without compromising causality learning.
• Anti-monotone properties are used to generalize causality and obtain complex causal paths.

Main Results:

• CauRuler was tested on a large medical database (3.5M visits, >15,000 variables).
• The rule set was reduced from 7,732 to 2,240 rules, identifying 46 causal relationships, generalized to 14.
• Obtained causal paths controlled an average of 906 confounder variables, yielding robust results validated by clinicians.

Conclusions:

• Causal relationships are vital for predicting health condition trajectories and preventing disease.
• CauRuler demonstrates efficiency and effectiveness in identifying known causal associations in medical data.
• The method's ability to control numerous confounders provides robust causal analysis for clinical applications.