Is there an information hierarchy among hemodynamic variables for early identification of occult hemorrhage?

Abstract

Objectives: In a clinical environment, the ability to provide assessments of a trauma patient's disease state as well as guidance on the potential impacts of interventions could inform decision making. We investigate the use of clinical data to explore models of disease progression, interventions, and outcomes. Methods: Time point measurements, including “clinically relevant data points,” such as temperature, heart rate, respiratory rate, pH value, Paco2, Pao2, HCO3, and “coagulation factor data points,” such as levels of fII, fV, fVII, fIX, fX, antithrombin, tissue factor pathway inhibitor, protein C, and activated protein C, of 859 patients were taken at intervals starting immediately after they entered the emergency department until, at most, 120 hours. Interventions (blood units, fresh frozen plasma units, platelet units, crystalloids, cryo units, colloid) provided to patients were recorded during nonoverlapping time intervals. Using these data, we have developed a linear model capable of estimating the combination of interventions that will be most effective in shifting a trauma patient to a desired phenotypic state (clinically relevant observables). In addition, we have developed a second model for estimating a patient's phenotypic state in the form of a frequently occurring combination of discrete measurements. Results: Our first model suggests that each unit of fresh frozen plasma provided will decrease active protein C by 2.28 (ng/mL) and increase pH value by 0.02068 (P= .000485). In addition, each unit of blood will increase fVII by 4.443 (% activity) (P = .005), and each unit of platelet will increase Paco2 by 5.246 (mm Hg) (P = .006). The model makes additional statistically significant predictions and can also accommodate other information such as demographics or category and severity of injuries. Using this methodology, we can predict the effects of manipulation of relative factor concentration on coagulation profiles and translate the optimal combinations into transfusion paradigms (eg, plasma-to-blood ratio). Identification of subgroups of patients and interventions will enable targeted therapies. Using our second model, we have discovered the 2 most frequent phenotypic states to be: low fII, fIX, fX, fVII, protein C, and high prothrombin time, which frequently occurs in patients before death, and high fII, fX, fVII, and high protein C, which frequently occurs in patients with a good prognosis. Understanding major states and how they transfer from one to another can reveal how trauma patients progress with medical treatments over time. Conclusions: We have developed a methodology to predict the effects of interventions on a trauma patient's phenotypic state. In addition, we have demonstrated how understanding of phenotypic states and their time evolution can reveal how trauma patients progress with medical treatments over time. Future work will focus on using models to (1) identify patient subgroups to design targeted interventions/therapies and (2) predict prognoses of newly admitted patients.