Sepsis Prediction at Emergency Department Triage Using Natural Language Processing: Retrospective Cohort Study

Area of Science:

• Computational medicine
• Clinical informatics
• Artificial intelligence in healthcare

Background:

• Sepsis presents diagnostic challenges in the emergency department (ED) despite its high mortality.
• Machine learning (ML) offers potential for early sepsis detection and intervention.

Purpose of the Study:

• To predict sepsis at ED triage using natural language processing (NLP) of nursing notes and clinical data.
• To evaluate ML model performance for sepsis prediction at various time points.

Main Methods:

• Retrospective cohort study of over 1 million ED encounters (2015-2021).
• Developed a decision tree-based ensemble model using vectorized triage notes and clinical data.
• Trained models on initial triage data and subsequently with hourly laboratory data.

Main Results:

• The time-of-triage model achieved an AUC of 0.94 and macro F1-score of 0.61.
• Sensitivity and specificity were 0.87 and 0.85, respectively, for sepsis prediction at triage.
• AUC improved to 0.97 with hourly laboratory data, predicting sepsis up to 12 hours prior to antibiotics.

Conclusions:

• Sepsis is accurately predictable at ED presentation using triage notes and clinical data.
• ML models, incorporating free-text data, can enable timely and reliable sepsis alerts.
• This approach can significantly improve early intervention for sepsis.