Introduction: Addressing the EMS Documentation Gap
In Emergency Medical Services (EMS), the pressure to provide life-saving care often conflicts with the need for meticulous documentation. Currently, most incident reports are compiled after the fact, relying on the memory of first responders. This "documentation gap" is significant; research indicates that approximately 13.6% of summary reports contain errors, ranging from missing data to misplaced information.
Automating this process is difficult due to the "noisy" nature of emergency scenes. Beyond literal acoustic noise like sirens and shouting, there is linguistic noise—non-linear narratives, specialized jargon, and a lack of punctuation in standard speech-to-text transcripts. GRACE (Generating Summary Reports Automatically for Cognitive Assistance) is an NLP-driven framework designed to bridge this gap by transforming real-time audio into structured medical reports.
The GRACE Architecture
The GRACE framework employs a multi-stage pipeline to process raw audio into actionable data.
The process begins with Speech-to-Text (STT) and Punctuation Restoration. Because STT often lacks the punctuation necessary for semantic clarity, GRACE uses a recurrent bidirectional neural network to "re-punctuate" the text. This is followed by Ensemble Concept Extraction, which utilizes three major clinical NLP tools: MetaMap, cTAKES, and CLAMP. By requiring a consensus (at least two tools agreeing on a concept), the system significantly reduces false positives caused by scene noise.
Heuristics and Semantic Inference
To handle the specific nuances of emergency dialogue, GRACE applies domain-specific heuristics. For instance, the "First Three Sentences" rule prioritizes early narrative information to identify the "Chief Complaint," as EMS providers typically lead with the primary issue.
Furthermore, the system incorporates Semantic Inference modules. It uses NegEx for negation detection—crucial for distinguishing between a patient having a symptom and denying one—and StanfordNER to link numerical values to specific vital signs, such as associating the number "99" with a patient's pulse.
Experimental Results
The system was evaluated using 8,000 real-world narratives and 119 simulated audio recordings, including versions with varying levels of ambient and artificial noise.
GRACE demonstrated high robustness, achieving an F1 score of 94% on noise-free audio and maintaining 78% on noisy audio. Specifically, the system reached an F1 score of 0.85 for Chief Complaint detection, outperforming standalone tools like MetaMap (0.67) and cTAKES (0.60). The adapted NegEx module also performed well, with an F1 score of 0.81.
Conclusion and Future Directions
GRACE represents a major advancement in cognitive assistance for emergency responders. While the system handles ambient noise and structured narratives effectively, future work will focus on improving performance in multi-patient scenarios and handling non-linear storytelling where information is provided out of order. Live deployment in high-stress environments remains the ultimate goal for validating the system's real-world utility.
Summary updated