Hemorrhagic shock occurs frequently in natural and man-made disaster scenarios. To control bleeding and to provide necessary resuscitation, swift and accurate diagnosis and decision-making are required. Early recognition of bleeding and the need for targeted interventions could improve both survival and resource management, allowing the receiving hospital to prepare required blood, surgeons, or other resources in advance of patient arrival and to conserve valuable resources in those patients who are not bleeding. Resources can be saved through avoidance of over-triage, thereby reducing unnecessary air transport, unnecessary blood transfusions, and unnecessary evaluation with labs, X-rays, and computed tomography scans, which is important in all resource-constrained and austere environments.
To achieve the above goals, many diagnostic and predictive models have been proposed. These models use information collected from patients, such as pre-hospital vital signs (VS), injury mechanism, or information measured from other devices. They are useful and some of them have been used in practice. However, those methods have their drawbacks. First, they are mostly based on a paper-pencil type scoring system. Users need to collect all variables and either do manual calculation or input the numbers to a calculator, costing valuable time. Second, alternative methods require specially trained experts to use certain devices, such as ultrasound machines. Third, other methods depend on blood sampling and laboratory testing; such requirement increases the logistics for in-the-field deployment. Moreover, there is no systematic validation of those models in military-specific populations.
Advances in computing and sensoring techniques allow real-time high-fidelity VS data collection and processing. Non-expensive and non-invasive sensors with built-in computing processors could fully automate data collection and calculation without user input during the multi-tasking trauma patient resuscitation period. Transfusion prediction algorithms use the non-invasive sensor signals to convert these data into clinically relevant quantities that can be used for identifying patients with life-threatening bleeding.
To derive such algorithms, a feasible and practical approach was to use a subset of >22,000 trauma patient datasets to train an algorithmic “learner.” By “observing” and “inferring” from the dataset, the association between input variables and the outcomes could be learned. This machine learning approach is called supervised learning. Moreover, the trained algorithms need to be thoroughly tested in testing datasets that the algorithms have never seen before. In this way, what performance to expect when the algorithms are applied in the future would be known. Therefore, it is important to use these validation methods to train and test the algorithms or models, e.g., for transfusion prediction, before deploying them for healthcare.
This has particular relevance for military applications. Hemorrhage is the greatest threat to survival, the leading cause of death, and the most common cause of potentially preventable combat-related mortality. The Department of Defense has invested many resources into developing reliable transfusion prediction models andpractical usable tools based on intensive analysis of large data collections. Military medicine considers these approaches as the future way to develop combat casualty autonomous resuscitation and enhance real-time field decision-making.
Past work has focused on the assessment and treatment of major bleeding in the pre-hospital environment, typically in the form of field severity scoring systems. On the other hand, using new sensors that are non-expensive and non-invasive, advanced field-ready algorithms and instruments that can be built for combat medics offers advantages. Such study can enhance the ability to rapidly assess fluid resuscitation needs. Through large-scale big data modeling, evidence-based physiologic criteria for validation of transfusion use in a combat casualty can be identified and thoroughly tested.
This work was done by Peter Hu, PhD; Shiming Yang, PhD; and Colin Mackenzie, MD; University of Maryland School of Medicine for the Air Force Research Laboratory. AFRL-0261
This Brief includes a Technical Support Package (TSP).
Validation of Automated Prediction of Blood Product Needs Algorithm Processing Continuous Non-Invasive Vital Signs Streams (ONPOINT4)
(reference AFRL-0261) is currently available for download from the TSP library.
Don't have an account? Sign up here.