Background: Injuries are a major health concern around the world, causing over 5 million deaths per year. Trauma accounts for more fatalities than HIV, malaria, and tuberculosis combined. Early, accurate, and transparent diagnoses of traumatic injuries are essential in this safety critical environment. Unfortunately, highly trained medical staff and sophisticated diagnostic equipment are often unavailable at the site of injury (e.g., low income countries, combat zones, and post-natural disaster areas).
Goal: An automated system for assisting first responders in diagnosing traumatic injury is needed. The availability of electronic health records offer great promise for correlating basic observable patient features such as Glasgow Coma Scores, vitals and injury mechanism, but the sheer volume of trauma patient data makes it virtually impossible to manually parse and glean useful information. Machine learning techniques can efficiently mine these data sets to pull in correlations and patterns which can then be exploited. Unfortunately, machine learning algorithms can often produce output which is opaque and/or confusing to the user. We propose an inexpensive injury prediction tool using sparse measurements capable of being taken from minimally skilled trauma responders and presenting the predicted injuries to the user, where those predictions can be further interrogated to provide a transparent, human-interpretable rationalization.
Methods: To counter the li
mited number of trained medical professionals and the low availability of cutting-edge diagnostic equipment in the aforementioned scenarios, we propose to leverage rich EHR in order to develop a useful traumatic injury screening tool. Given there are multiple injuries and each patient may have multiple injuries, we treat this as a multiclass and multilabel classification problem. Specifically, we will leverage the non-linear classifying capability of decision trees to diagnose trauma victims using only a small number of admissible metrics (e.g., heart rate, breathing rate, Glasgow coma scores, etc.). Furthermore, decision trees can deliver a transparent and interpretable summary of the classification decision to the user.
Results: From the Trauma Registry maintained by the R. Adams Cowley Shock Trauma Center of the University of Maryland, we gained electronic health records for 2,651 visits by 2,643 distinct patients. Due to the low number of instances of most injuries, we considered the top 34. The admission metrics serves as features in the classification problem and act as the input: Injury Location, Injury Type, Protective Equipment, Abbreviated Injury Scale (AIS) Injuries (which describes the specific injuries suffered by each patient) Glasgow Coma Scale (GCS), Region of injury numerical injury index. Across the 34 injuries, we produced the following summary statistics: Sensitivity = 0.80, Specificity = 0.98, Precision = 0.77, F1 = .0.79.
Discussion: The applications of such a prediction tool are far-reaching. For example, we have developed an injury diagnosis Android application leveraging our learned decision tree models. First responders, combat medics, and disaster relief medical staff could be equipped with such an app to assist them in making informed diagnoses and emergency treatment for victims of trauma, without extensive training or resources.
DECISION SUPPORT & HOSPITAL MONITORING
Author: Jeff Druce
Coauthor(s): Max Metzger – Charles River Analytics Nidhi Gupta – Charles River Analytics Rishi Kundi – Division of Vascular Surgery, University of Maryland School of Medicine
Status: Work In Progress
Funding Acknowledgment: The author wishes to acknowledge the contributions and support of Dr. Gary Gilbert, Mr. Carl Manemeit, and Ms. Rebecca Lee of the Telemedicine and of the Telemedicine & Advanced Technology Research Center (TATRC). The authors additionally wish to acknowledge the contributions of Dr. Todd Rasmussen of the R. Adams Cowley Shock Trauma Center of the University of Maryland Medical Center as well as Dr. Thomas Scalea, the Physician-in-Chief of the Center. Additional thanks are given to Dr. Rajabrata Sarkar, the Chief of the Division of Vascular Surgery at the University of Maryland School of Medicine. This work was funded by the US Army Medical Research and Materiel Command (USAMRMC) Telemedicine & Advanced Technology Research Center (TATRC).