Automated prediction of early blood transfusion and mortality in trauma patients

Background: Prediction of blood transfusion needs and mortality for trauma patients in near real time is an unrealized goal. We hypothesized that analysis of pulse oximeter signals could predict blood transfusion and mortality as accurately as conventional vital signs (VSs). Methods: Continuous VS data were recorded for direct admission trauma patients with abnormal prehospital shock index (SI = heart rate [HR] / systolic blood pressure) greater than 0.62. Predictions of transfusion during the first 24 hours and in-hospital mortality using logistical regression models were compared with DeLong's method for areas under receiver operating characteristic curves (AUROCs) to determine the optimal combinations of prehospital SI and HR, continuous photoplethysmographic (PPG), oxygen saturation (SpO2), and HR-related features. Results: We enrolled 556 patients; 37 received blood within 24 hours; 7 received more than 4 U of red blood cells in less than 4 hours or "massive transfusion" (MT); and 9 died. The first 15 minutes of VS signals, including prehospital HR plus continuous PPG, and SpO2 HR signal analysis best predicted transfusion at 1 hour to 3 hours, MT, and mortality (AUROC, 0.83; p < 0.03) and no differently (p = 0.32) from a model including blood pressure. Predictions of transfusion based on the first 15 minutes of data were no different using 30 minutes to 60 minutes of data collection. SI plus PPG and SpO2 signal analysis (AUROC, 0.82) predicted 1-hour to 3-hour transfusion, MT, and mortality no differently from pulse oximeter signals alone. Conclusion: Pulse oximeter features collected in the first 15 minutes of our trauma patient resuscitation cohort, without user input, predicted early MT and mortality in the critical first hours of care better than the currently used VS such as combinations of HR and systolic blood pressure or prehospital SI alone. Level of Evidence: Therapeutic/prognostic study, level II.