Predicting Critical Outcomes in Suspected Cardiopulmonary Emergencies Using Dispatch Narratives: Temporal Validation...

Area of Science:

• Emergency medicine
• Data science
• Artificial intelligence

Background:

• Early risk stratification is crucial for patients with acute cardiopulmonary symptoms in emergency medical services (EMS).
• Prehospital decision-making often relies on limited structured information, overlooking potential clinical signals in free-text dispatch narratives.

Purpose of the Study:

• To develop and validate a natural language processing (NLP)-assisted machine learning framework for early risk stratification using free-text EMS dispatch narratives.
• To assess the added value of NLP-derived features beyond conventional structured dispatch data.

Main Methods:

• A retrospective cohort study of 38,523 EMS dispatch records (2021-2025) for suspected cardiopulmonary symptoms.
• A multimodal machine learning framework combining NLP-processed character-level n-grams from narratives with structured variables (age, sex, call time).
• Temporal validation using data from 2021-2024 for development and 2025 for testing.

Main Results:

• The NLP-enhanced model achieved an AUROC of 0.803, significantly improving upon the structured baseline model (AUROC 0.681).
• The top 10% of predicted high-risk cases represented a 3.2-fold enrichment of critical outcomes.
• Model performance was consistent across subgroups, demonstrating acceptable calibration and net benefit.

Conclusions:

• Free-text EMS dispatch narratives contain valuable information for early risk stratification in cardiopulmonary emergencies.
• Integrating narrative features into structured models can enhance EMS dispatch systems and prehospital decision-making.
• Further external validation and prospective studies are recommended.