Leveraging Expert Knowledge and Causal Structure Learning to Build Parsimonious Models of Acute Brain Dysfunction in...

Area of Science:

• Biomedical Informatics
• Clinical Decision Support
• Artificial Intelligence in Medicine

Background:

• Machine learning (ML) in clinical decision support systems faces challenges with transparency and robustness.
• Causal structure learning (CSL) combined with expert knowledge can enhance ML model interpretability and clinical alignment.
• Acquired acute brain dysfunction (ABD) in the pediatric intensive care unit (PICU) requires robust predictive modeling.

Purpose of the Study:

• To integrate clinician expertise with CSL algorithms to identify causal drivers of acquired ABD in the PICU.
• To develop parsimonious predictive models for ABD without significant performance loss.
• To enhance the transparency and clinical relevance of ML-based decision support.

Main Methods:

• Analysis of 18,568 PICU encounters from a major medical center (2010-2022).
• Elicitation of expert knowledge from four clinicians to construct a consensus directed acyclic graph (DAG).
• Application of CSL algorithms (GOLEM, PC-MB) to enrich the expert-derived DAG and development of XGBoost predictive models.

Main Results:

• Clinician consensus identified 16 potential causal biomarkers for ABD with acceptable inter-rater reliability (Fleiss' Kappa = 0.62).
• PC-MB algorithm showed 78% concordance with expert consensus, identifying seven additional potential causal biomarkers.
• XGBoost models using 14 biomarkers from the intersection of expert consensus and PC-MB achieved an AUPRC of 0.79, comparable to a model using all 45 biomarkers (AUPRC = 0.81).

Conclusions:

• Combining clinical expertise with CSL effectively identifies plausible causal drivers of acquired ABD.
• This integrated approach facilitates the development of parsimonious and clinically aligned predictive models for ABD in the PICU.
• The study demonstrates a viable strategy for improving the transparency and robustness of ML in clinical decision support.