Machine Learning Prediction of Blood Loss from Arterial Waveform Feature Analysis in a Nonhuman Primate Hemorrhage Model

Hemorrhage is the leading cause of preventable death in trauma, yet traditional vital signs can fail to detect early blood loss due to physiological compensatory mechanisms. Objective, continuous monitoring of hemorrhage progression remains an unmet clinical need. This study developed and validated a machine learning model of percent estimated blood loss in a nonhuman primate hemorrhage model. Adult male baboons underwent controlled stepwise hemorrhage to approximately 25% blood loss while high-fidelity arterial waveform and blood loss data were captured. Morphological waveform features were extracted, and a Gradient Boosting model was developed using leave-one-subject-out cross-validation. A 10 s window with 10 features achieved a mean coefficient of determination of 0.81 and mean absolute error of 2.5%. On three independently held-out subjects, per-subject coefficient of determination ranged from 0.75 to 0.94 while mean absolute error ranged from 1.4% to 3.3%. Notably, percent estimated blood loss detected hemorrhage a median of 8.1 min before sustained blood pressure decline, at approximately 7% blood loss while systolic pressure remained stable. These findings indicate that arterial waveform morphology can enable accurate, continuous blood loss estimation and may provide early warning before hemorrhage becomes detectable through conventional vital sign monitoring in this preclinical model.