JMIR Publications

ABSTRACT

Background:

Almost 40% of children admitted to the pediatric intensive care unit (PICU) acquire secondary brain injuries or related morbidities, encompassing a wide range of diagnoses, each with different causes and treatments. To reduce related harms, clinicians need to identify at-risk patients as early as possible, which can be enabled by predictive models. However, the adoption of machine learning remains limited by concerns about model transparency and robustness. Causal structure learning (CSL) combined with expert knowledge may address these concerns by identifying potentially causal predictors, enabling more interpretable and clinically aligned models.

Objective:

In this study, we explore whether integrating clinician expertise with CSL algorithms may enable us: (1) To identify plausible causal drivers of acquired acute brain dysfunction (ABD) in the pediatric intensive care unit (PICU) and (2) Develop predictive models for ABD with a limited set of biomarkers without substantial loss in performance.

Methods:

We analyzed 18,568 PICU encounters from the University of Pittsburgh Medical Center Children’s Hospital (2010–2022) and elicited knowledge from experienced clinicians. Encounters with acquired ABD were defined using the validated ABD computable phenotype. Expert knowledge was elicited from four clinicians through iterative interviews to construct a consensus directed acyclic graph (DAG). Two CSL algorithms, GOLEM and PC-MB, were applied to enrich the clinician’s consensus DAG. Using multiple variations of the enriched DAGs, XGBoost models were trained using biomarkers identified as potential causes of acquired ABD; these were evaluated primarily by area under the precision-recall curve (AUPRC) due to the class imbalance in our dataset.

Results:

The clinician consensus DAG achieved acceptable inter-rater reliability (Fleiss’ Kappa = 0.62) after two rounds of interviews and identified 16 biomarkers as potential causes of acquired ABD. The PC-MB algorithm showed 78% concordance with clinician consensus, while GOLEM showed 46%. Together, the CSL algorithms identified seven biomarkers as potential causes that were not included in the clinician’s DAG: blood urea nitrogen, creatinine, dobutamine, glucose, potassium, PTT, SpO2. Predictive models trained on the intersection of clinician consensus and PC-MB DAGs achieved an AUPRC of 0.79 (95% CI: 0.75–0.82) using only 14 biomarkers, compared to an AUPRC of 0.81 (95% CI: 0.78–0.84) for the control model using all 45 biomarkers. When restricted to vitals and laboratory results alone, the best-performing model achieved an AUPRC of 0.77 (95% CI: 0.74–0.81).

Conclusions:

Combining clinical expertise with causal structure learning enables the identification of causal hypotheses consistent with the clinical understanding of the participating clinicians and the development of more parsimonious predictive models for acquired ABD in the PICU.