Poster #73: Kayra Saçan

Predicting Survival in Hemorrhagic Shock using Machine Learning​

Saçan, K., Shoshany, D., & İzzetoğlu, K., PhD. Drexel University, College of Engineering and Computing, School of Computer & Information Sciences¹, School of Biomedical Engineering & Science².

Hemorrhagic shock accounts for more than 80% of preventable post-injury deaths, making early detection critical for improving patient outcomes. The current standard of care for monitoring relies primarily on systemic vital signs, such as heart rate (HR) and blood pressure (BP). However, changes in vital signs can be nonspecific and may only become apparent during later stages of blood loss. Near-infrared spectroscopy (NIRS) and diffuse correlation spectroscopy (DCS) are emerging optical technologies that enable continuous, non-invasive measurement of cerebral oxygenation and blood flow. Because compensatory physiological responses during early hemorrhage prioritize perfusion of vital organs, including the brain, measurements of cerebral hemodynamics may reveal early physiological changes associated with hemorrhage and may help predict survival outcomes. In this study, eight machine learning models were developed and evaluated to predict survival outcomes using multimodal physiological data from swine models of hemorrhagic shock (N=20). Features included systemic vital signs as well as cerebral measurements obtained using NIRS and DCS. Feature importance analysis identified relative blood flow index (rBFI), a cerebral index, as the strongest predictor of mortality, followed by systolic blood pressure (SBP), a systemic index. A multimodel soft-voting ensemble using the five most important features achieved the highest performance, with accuracies of 77.5% using cerebral features alone, 88.2% using systemic features alone, and 90.2% using the combined feature set. These preliminary findings demonstrate the potential of integrating NIRS- and DCS-derived cerebral hemodynamic measures with traditional systemic vital signs to predict survival outcomes. Cerebral optical measurements provide complementary information to traditional vital signs, while combining both modalities yields the strongest predictive performance. Future work will expand the dataset, optimize model performance, and develop an interpretable risk score capable of supporting early clinical decision-making in trauma care.