A novel method for causal structure discovery from EHR data and its application to type-2 diabetes mellitus

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Clinical Decision Support

Background:

• AI clinical decision support models utilize numerous predictors, but associative relationships limit their robustness for intervention planning.
• Traditional causal discovery methods (clinical trials, biochemical pathways) are resource-intensive and struggle with complex, large-scale data.
• Existing computational causal structure discovery (CSD) methods are not optimized for electronic health records (EHR) data.

Purpose of the Study:

• To present a novel CSD method specifically adapted for the unique characteristics of EHR data.
• To improve the reliability and scalability of causal relationship discovery for precision medicine.
• To enhance the suitability of CSD for clinical decision support applications.

Main Methods:

• Developed a new CSD methodology tailored to the complexities of EHR data.
• Applied the method to a large EHR dataset from Mayo Clinic for development.
• Validated the method on an independent EHR dataset from M Health Fairview.

Main Results:

• The proposed CSD method demonstrated very high recall (0.95).
• Achieved substantially higher precision compared to general-purpose CSD methods (0.84 vs. 0.29 and 0.55).
• High overlap (81%) in extracted causal relationships between development and validation cohorts indicates robustness.

Conclusions:

• The EHR-tailored CSD method is more effective for clinical decision support than general-purpose approaches.
• The findings support the use of this method for discovering causal relationships in EHR data for precision treatment.
• The high overlap across cohorts suggests the method's generalizability and reliability.