The goal of this grant is to 1) form a computer-aided system to improve decision-making for traumatic pelvic injuries and 2) support research on a number of data integration topics in computer science. Trauma is the leading cause of death for Americans under the age of 45. Traumatic pelvic injuries can be fatal due to severe hemorrhage, and care givers treating these injuries need to consider several types of data, including biomedical signals, images, trauma scores, laboratory results, diagnosis/treatment, injury specifics, and demographics during the decision making process. Integrating simple types of data such as lab results and demographics is not easy, but the decision-making process shows its true complexity when trying to integrate more complex types of patient data such as biomedical signals and images.
Intellectual Merit The project is challenging in the following aspects: o It constructs a traumatic pelvic injury database that includes all relevant biomedical signals/images, trauma scores, lab results, diagnosis, treatment, demographics, and injury specifics for each patient. This database will have two significant advantages over existing databases: 1) it will contain not only patient demographics and trauma scores but also time-series (signals) of physiological measures and images; and 2) in the new database, instead of including only raw data, patient information is processed and transformed into a set of features that can be directly used for decision making. o A variety of novel biomedical signal and image processing methods will be formed to extract relevant features. These computational methods will include both the improved versions of computational methods in signal and image processing (e.g., segmentation techniques for CT images), and feature extraction methods for specific signals and images (e.g., defining the total area of the pelvic ring captured from CT as a feature). o The project constructs a rule database where all derived features for patients in the feature database are analyzed with outcomes, resulting in a set of rules to describe logical relationships among the input features and resulting outcomes/recommendations, using non-linear classification and regression tree. The project is novel in its rule validation; besides using typical statistical methods such as cross-validation and measures of sensitivity and specificity used in existing systems, a new statistical framework based on computational learning theory will be used to allow a more comprehensive comparison of the new system with other methods such as neural networks and Bayesian classifiers.
Broader Impacts This project brings together computer scientists with trauma experts and will produce a system that can be replicated at other hospital systems. This methodology can be used for other types of trauma cases, such as brain injuries. Educationally, project results will be included in regular seminars to teach healthcare providers across the spectrum of the trauma care the latest techniques used in trauma care. The PI will align the research project with undergraduate and graduate research and outreach activities managed by the University of North Carolina at Charlotte''s Diversity in IT Institute, whose mission is to increase enrollment and retention of women and underrepresented groups within IT with a focus on facilitating graduate and undergraduate interdisciplinary programs, and two NSF-funded programs housed within the institute: 1) The Students & Technology in Academia, Research, and Service Alliance: A Southeastern Partnership for Broadening Participation in Computing, and 2) Computing Research for Undergraduates.
Traumatic pelvic injuries are among the most severe injuries that can be suffered by a trauma patient. The mortality rate is increased by the risk of complications; in particular, hemorrhage and laceration of the surrounding soft tissue and neural and vascular structures. Prompt and appropriate treatment of pelvic injury is therefore crucial to patient survival. A computer assisted decision making system capable of rapidly analyzing large volumes of patient information to generate accurate treatment recommendations and outcome predictions has the potential improve both patient care and resource utilization. Objective: The main objective of this project was to construct a computer-assisted decision-making system for pelvic trauma to process all available patient data in order to generate useful recommendations and predictions at every stage of care. This data includes common measurements such as GCS and ISS, medical images (CT, X-Ray etc), physiological signals (such as ECG and BP), demographic information, lab results, and various other sources. The main goals of the project were: Form a database of traumatic pelvic injuries that contain demographics, trauma scores, pelvic images (e.g. CT or X-ray images), and physiological signals (e.g. blood pressure). Design of signal processing methods to analyze signals including ECG and blood pressure to extract diagnostically important features. Design machine learning methods to consider extracted features and form diagnostic as well as therapeutic predictions / recommendations for care givers. Methods and Results: A large database consisting of about 60,000 Traumatic Pelvic Injury images were formed. One of the main focuses of the project was on processing of CT and X-Ray images, in particular to detect fracture and hemorrhage, and quantitatively evaluate important displacement measures. X-Ray images are first segmented using a deformable model approach, incorporating splines to regulate deformation in identifying key pelvic bone structures. This algorithm is also unique in its ability to segment all structures in one pass via a hierarchical automatic initialization process, removing the need for user interaction. The detected structures are then analyzed for fracture, using methods such as Discrete Wavelet Transform (DWT) to detect changes in texture and discontinuities in bone outline. Displacement measurements are also calculated - for example, the width of the pubis gap. CT scans can offer better visualization of hemorrhage, if this is suspected based on the initial X-Ray. A multi-stage algorithm is used to identify bone structures across multiple CT slices. First, the desired region is extracted and filtering is applied to highlight bone. Segmentation is performed using a snake algorithm initialized with seed-growing. This method has proven accurate in detecting edges and offers good separation between nearby bones. The method also detects both bone fracture and the presence and size of hemorrhage. In signal processing stage, ECG and other signals such as BP, are analyzed using methods such as DWT to extract characteristic features that can distinguish healthy from diseased cases. In particular, the features developed in this project were designed to detect the presence and severity of hemorrhage, which is the main cause of death and long-term complications for most types of injuries. Once key features are extracted and combined with other information, such as physiological and demographic information, our machine learning method, which is an improved version of ECOC, is used to create the output recommendations and predictions. These predictions allow physicians to evaluate the reasoning behind them and make a more informed decision on whether they should be adopted or disregarded. Some of the defined features prove to be very significant in clinical applications; e.g. X-Ray displacement measures appear to be particularly significant in predicting a patient's length of stay in ICU after pelvic trauma. The results of the project were published in international conferences and journals.
