Explaining alerts from a pediatric risk prediction model using clinical text.

Area of Science:

• Pediatric critical care medicine
• Clinical informatics
• Artificial intelligence in healthcare

Background:

• Risk prediction models are crucial for identifying at-risk pediatric patients.
• Timely interventions are essential for preventing clinical deterioration.
• Existing models often lack explainability for alert triggers.

Purpose of the Study:

• To develop a novel explainer algorithm for risk prediction alerts.
• To generate text-based explanations from patient clinical notes.
• To improve the interpretability of the pediatric Calculated Assessment of Risk and Triage (pCART) model.

Main Methods:

• Retrospective study of 39,406 pediatric patient admissions.
• Utilized the validated pCART risk prediction model.
• Trained a transformer model with label-aware attention on clinical notes preceding alerts.
• Split data into derivation, validation, and test sets for performance evaluation.

Main Results:

• The pCART Explainer algorithm demonstrated strong performance in discriminating at-risk alerts (c-statistic 0.805).
• Explanations highlighted clinically significant phrases like "rapid breathing" and "fall risk."
• The algorithm showed excellent face validity, confirming clinical relevance.

Conclusions:

• Developed pCART Explainer, a novel algorithm for explaining deterioration alerts.
• The algorithm provides medically relevant context by highlighting key phrases in clinical notes.
• pCART Explainer can enhance clinician situational awareness and guide decision-making.