Words to Live by: Leveraging Natural Language Processing and Machine Learning to Enhance Prehospital Triage Algorithms

About This Grant

Project Summary Abstract Over 1 million EMS clinicians from over 23,000 agencies must determine illness severity and care needs for the 25 million US patients transported annually to emergency departments. Treatment decisions based on these assessments—or 'triage’, such as whether to administer a lifesaving intervention (LSI) or transport to a Level 1 trauma center—proves challenging: Under- and over-estimation of illness severity and care needs occur at rates of 11%-72% and 10%-48%, respectively, which annually leads to thousands of excess deaths and costs the health care system hundreds of millions of dollars. Prehospital clinicians struggle to make accurate triage decisions because they have little available diagnostic information or time for analysis. Prehospital clinicians could benefit if diagnostic information were synthesized via an algorithmic tool to support triage decisions. To this end, prehospital clinicians record several lexical observations regarding patient condition and care needs, including clinician impressions and 9-1-1 telecommunicator notes. These observations mix diagnostic cues (e.g., anatomic injury patterns; possible illness etiology) with provider intuition, which is itself predictive of illness severity and care needs. Words within prehospital lexical observations may inform triage decisions if incorporated into an algorithmic decision tool. This project will involve the first comprehensive test of whether and how lexical prehospital information can be leveraged via natural language processing (NLP) and machine learning (ML) to create triage algorithms. NLP and ML prediction models trained on free-text prehospital clinician impressions will be used to predict illness severity (e.g., Injury Severity Score; mortality; hospital length-of-stay) and administration of prehospital LSI (e.g., intubation; defibrillation; tourniquet). To promote generalizability, models will be built in three large data sets totaling over 12 million prehospital cases; these cohorts vary in transport mode (i.e., ground; air), medical condition (i.e., trauma; non-trauma) and free-text format (i.e., 2-3 word clinician impressions; 4-5 sentence anatomic descriptions; 9-1-1 call notes). Multiple state-of-the-art NLP and ML approaches (e.g., ensemble models using bag-of-words frequencies; transformer models with pretrained embeddings) will be used to balance clinical interpretability and predictive sophistication. Highly predictive ML models could shape triage protocols: Notes from 9-1-1 calls could be fed into an ML prediction model to inform delivery of appropriate support and resources to a patient’s side (e.g., advanced clinicians; blood product); voice-to-text recordings of clinician impressions made upon patient encounter could likewise be leveraged to determine the need for lifesaving care. This work will set the stage for prospective collection of prehospital lexical data, as well as videos of patient encounters in the field, leveraging voice-to-text translation and computer vision-generated scene descriptions to translate these data sources into real-time decision support tools for prehospital triage.