Author(s): Catherine Clive, Spencer Boris, Jaden Searle, Lydia Peterson, Greg Hooke, Bram Overmeer
Mentor(s): Brandon Westover, Haoqi Sun
Institution BYU
Conducting medical studies can often be time-intensive, costly, and pose significant risks to patients involved. Retrospective studies utilizing electronic health records (EHRs) present a viable alternative, offering a resource-efficient and unintrusive method for data collection and analysis. In this study, we focused on the investigation of EHRs in the context of anoxic brain injury and the withdrawal of life-sustaining treatment (WLST), with the aim of developing a tool to identify patients who receive these diagnoses. We employed artificial intelligence (AI) to construct two distinct logistic regression models designed to classify EHRs based on specified criteria. The dataset used for training and validation was sourced from two affiliated institutions—Brigham and Women’s Hospital and Massachusetts General Hospital—as well as an independent institution, Beth Israel Deaconess Medical Center. We utilized the patient notes included in their records, as well as ICD codes and CPT codes: codes associated with diagnoses, and scan orders respectively. Both models demonstrated an accuracy exceeding 95%, indicating an extremely reliable ability to distinguish between patient groups with each diagnosis. The developed tool provides a novel approach for investigating the self-fulfilling prophecy associated with WLST. Specifically, it addresses the potential bias in survival statistics, which may be influenced by the inclusion of patients who undergo WLST. This feedback loop can result in decisions to withdraw life support being based on survival data that inherently includes patients whose treatment was withdrawn, thus perpetuating a cycle of biased outcomes. The tool's high accuracy enables robust analysis and could contribute to more nuanced understanding and decision-making in critical care practices.