Machine learning for prehospital prediction of trauma-induced coagulopathy: a national evaluation of the potential impact on tranexamic acid administration.

BACKGROUND

Trauma-induced coagulopathy (TIC) is characterized in part by hyperfibrinolysis. Inhibiting fibrinolysis with tranexamic acid (TXA) decreases mortality in trauma patients, with greater benefit the earlier it is given. However, identifying which patients could benefit has been the subject of controversy. Machine learning models that predict TIC early after injury may aid clinician decision-making. The aim of this study was to identify the proportion of patients with different risks of developing TIC based on TIC-Bayesian network (BN) who received TXA within 1 hour of injury.

METHODS

This retrospective cohort study included all prospectively-collected adult patients in the UK Trauma Audit Research Network database, between 2015 and 2019, who received hemorrhage control interventions or blood transfusions. Patients were risk-stratified using TIC-BN. Patients were categorized by whether they received TXA ≤1 hour from injury, 1-3 hours from injury, >3 hours from injury, or did not receive it.

RESULTS

Of 27,272 included patients, median age was 51 years (IQR 32-71), 68% were male, 86% suffered blunt injury, and median Injury Severity Score was 13 (IQR 9-25). Overall, 5571 (20%) patients received TXA ≤1 hour from injury, including 368/3,403 (11%) with very low risk, 494/4,764 (10%) low risk, 1,805/9,408 (19%) medium risk, 2,003/8,173 (25%) high risk, and 901/1,524 (59%) very high risk of TIC (χ2 for trend: p<0.001). Of those with medium or higher TIC risk, 75% did not receive TXA ≤1 hour.

CONCLUSIONS

Earlier identification of patients that may benefit from TXA using a machine learning risk prediction model may allow more tailored and prompt treatment for patients at risk of TIC.

LEVEL OF EVIDENCE AND TYPE OF STUDY: Level III, Retrospective Cohort.