Blood Loss Severity Prediction using Game Theoretic Based Feature Selection Abolfazl Razi 1 , Fatemeh Afghah 2 , Ashwin Belle 3 , Kevin Ward 3 and Kayvan Najarian 4 , Abstract— Detection of hypovolemia in the early stages of hemorrhage is an important but unsolved problem in medicine. Many preventable deaths amongst critically injured patients happen due to delayed treatment of uncontrolled hemorrhage. Using a database of physiological signals collected during simulated hemorrhage on human subjects, our research applies a variety of signal processing techniques to extract a multi- tude of features that enable the prediction of the severity of hemorrhage. In this study, a novel feature selection method based on coalition game theory has been proposed which helps identify the most valuable features and thereby reduce the size of the feature space. This reduction in feature space not only improves the efficiency, but also improves the prediction accuracy and reliability of the developed model. This feature selection algorithm is independent of the underlying classifi- cation method and can be combined with any classification method based on the employed data. The proposed feature selection method significantly enhances the prediction accuracy by optimally selecting the features compared to the state of the art. I. I NTRODUCTION Trauma induced hemorrhage has been a major factor in preventable deaths, both in civilian and combat settings [1], [2]. External or internal Hemorrhage when untreated can quickly lead to hypovolemia followed by hemorrhagic shock. Factors such as the types of treatment and how quickly it is administered can hinge the overall chances for survival of trauma patients. In combat settings nearly 20% of the trauma patients die even before reaching a treatment facility, of which nearly 50% of the deaths are caused by hemorrhage [3], [4]. Hence rapid assessment of the severity of hemor- rhage as well as accurate triage decision and treatment is crucial for the survival of such patients. However, assessment of hypovolemia and the severity of blood loss in a patient can be a challenging problem, especially considering the time critical nature of hemorrhage. Therefore the ability to detect changes in blood volume and to be able to predict the severity of blood loss can be very important in providing early and successful intervention. There is a need for such a system since available physiological signals from the patient such as heart rate, oxygen saturation, arterial blood pressure and arterial hemoglobin does not reveal any early signs of hemorrhage until the onset of cardiovascular decompensation 1 A. Razi is with the Department of Electrical and Computer Engineering, Duke University, Durham, NC 27708 abolfazl.razi@duke.edu 2 F. Afghah is with the department of Electrical and Computer Engi- neering, North Carolina A&T State University, Greensboro, NC 27410 fafghah@ncat.edu 3 A. Belle and K. Ward are with the Department of Emergency Medicine and Michigan Center for Integrative Research in Critical Care, University of Michigan, Ann Arbor, MI 48109 {bellea, keward}@umich.edu 4 K. Najarian is with Department of Computational Medicine and Bioinformatics, and Michigan Center for Integrative Research in Critical Care, University of Michigan, Ann Arbor, MI 48109 kayvan@med.umich.edu [5], by which time it could be too late. In our previous work we successfully developed a variety of signal processing and machine learning based systems which extracted fea- tures from heart rate variability (HRV) [6], morphology of electrocardiogram (ECG) [7], and from other physiological signals [8]. These systems utilized the variety of features and bio-markers that were extracted from these signals to predict the severity of hemorrhage. However, as the size of the features space being extracted grew larger, it became obvious that the size and quality of this feature set can highly affect the efficiency and performance of machine learning algorithms and its predictive accuracy. Hence, the following study proposes a novel game theoretic based feature selection approach, which improves the efficiency and accuracy of hemorrhagic shock prediction. Using large feature sets imposes some restrictions during storage, search and classification steps. The possibility of redundant information as well non-informative features tends to cause a whole host of problems during the classification stage such as inefficiency, over fitting, reduction in accuracy etc. To overcome these issues and improve classification outcomes, it is important to reduce the feature space to a more concise and relevant set of features. Different feature selection algorithms have been studied in the literature [9]–[11] to reduce the data dimensions and recognize the irrelevant and redundant features. The irrelevant features do not have any useful information related to the target, while the redundant features do not provide any more information than the features, which have been already selected. These mechanisms improve the prediction algorithm performance and also reduce the required storage. In general, these meth- ods are divided into three categories of embedded, filtering, and wrapper methods. In embedded methods, the feature selection is not per- formed implicitly, rather the classification model is such that the contribution of irrelevant features become limited. For instance, in a fully Baysian RVM method, the irrelevant features does not contribute to the classification due to the vanishing corresponding coefficients in the model [12], [13]. However, these method are not desired in applications that the explicit list of active features are desired. For instance, in a genomic data analysis, providing the list of contributing genes is an essential requirement. The wrapper methods are based on the learning algorithms, where the classifier is re- trained for any new data sets. Although this method results in a good performance, intensive required computations and the risk of over fitting limit the application of this approach in large datasets. Moreover, these methods are sensitive to the classification method and should be performed for each new classification method [14], [15]. Filtering methods, on the other hand, do not use a learning mechanism for feature 978-1-4799-2131-7/14/$31.00 ©2014 IEEE 776