Advancements in civilian casualty care can reduce mortality from trauma, particularly those due to airway compromise or prehospital haemorrhage. Timely delivery of lifesaving interventions (LSIs) is crucial but challenging for prehospital medics, who must make rapid decisions under constrained conditions with limited experience and unreliable triage guidelines.
Machine learning (ML) offers a promising solution by analysing high-dimensional physiological waveform data for predicting mortality, LSI need, and clinical deterioration. Features like heart rate variability and photoplethysmography enhance predictive accuracy.
In a recent study, researchers developed an ML model using brief physiological waveform data to predict the need for LSIs in individual trauma patients. The model was evaluated on standard ML performance metrics and compared to national triage benchmarks for overtriage and undertriage rates, with the hypothesis that it would perform comparably to or better than current methods.
The study analysed data from critically ill trauma patients transported by a major critical care air transport system in Pennsylvania and nearby states between January 1, 2018, and November 18, 2021. Patients were included if prehospital clinicians identified them as trauma cases during scene runs. Metrics were derived from physiological waveform signals and vital sign patterns recorded during the first 15 minutes after patient care and transport began.
Administration of an LSI occurred within a 2-minute care window. An ensemble machine learning model was used to predict LSI based on physiological data from the preceding 2-minute epoch.
The study included 2,809 participants (mean age 47.7 years; 70.5% men) with 15,088 two-minute physiological data epochs, 6.0% of which involved an LSI. The machine learning model showed good overall performance in predicting LSI, with an AUC of 0.81. It had high specificity (96%) and negative predictive value (95%) but lower sensitivity (27%) and positive predictive value (30%). The model performed equally well or better when predicting specific LSI types (e.g., airway intervention, blood transfusion), using data from up to 15 minutes before LSI, predicting only the first LSI per patient and across different injury mechanisms.
This study showed that continuous physiological waveform data collected early during prehospital care can accurately predict whether a trauma patient will receive an LSI within a 2-minute window during transport. The model had a 21.3% undertriage rate, higher than the national goal of 5%, and a 34.9% overtriage rate, which meets the national standard of 35% or less. It maintained accuracy for specific LSI types and up to 15 minutes before intervention.
These findings suggest important implications for prehospital triage, where current methods often rely on limited objective data and clinician intuition, leading to higher under- and overtriage rates than recommended. Compared to existing clinical models, this model performs favourably and provides time-specific predictions that could help prioritise patients for LSIs in challenging multi-casualty situations.
By automating detection of when a patient needs critical care, the model could ease the cognitive load on medics, allowing them to focus on the most severe or unclear cases. Since it uses routinely collected physiological data but analyses far more information than humans can process, the model offers a practical, transparent decision support tool that could integrate smoothly into prehospital workflows without disrupting medics’ routines.
Source: JAMA
Image Credit: iStock
