The following Great Innovative Idea is from Katie Henry, a current PhD student in computer science at Johns Hopkins University. In addition to the department, Henry is also part of the Malone Center for Engineering in Healthcare, the Institute for Computational Medicine, and the Center for Language and Speech Processing. Henry presented her poster, Automated in-patient monitoring in the ICU with application to septic shock prediction, at the CCC Symposium on Computing Research, May 9-10, 2016.
The Innovative Idea
Traditional approaches to disease prediction involve a panel of experts selecting a small set of clinically meaningful measurements and using these to tabulate a score. While useful, these scores are limited because they require manual definition and testing for each new disease and are limited to features that are easy for a human to compute in their checklist. Instead, we can use machine learning techniques to automatically learn features from routinely collected data in electronic health records (EHRs) that predict which patients are at highest risk of developing a given adverse-event. As a test case, we developed TREWScore, a targeted real-time early warning score for septic shock, a whole body infection that causes organ dysfunction and dangerously low blood pressure. While best practices for treatment are still under debate, there is consensus that early intervention is critical. Current approaches to identify septic shock use checklists to detect septic shock at the actual onset of shock (systolic blood pressure < 90 mmHg); however, TREWScore was able to identify patients with a median 28 hours prior to septic shock onset at a sensitivity of 0.85 and corresponding specificity of 0.67. Additionally, over two-thirds of patients were identified prior to any sepsis-related organ dysfunction.
Septic shock is the 11th leading cause of death in the United States and with $15.4 billion in annual health care costs, it has the highest associated added costs of any ICU condition. While the true impact of a septic shock prediction score like TREWScore has to be validated in a prospective study, there are several potential impacts that it could have. First, by allowing clinicians to identify at-risk patients earlier, they may be able to start treatment earlier, which has been shown to decrease mortality and morbidity. Second, septic shock has been referred to as a “pharmaceutical graveyard” in part because the large heterogeneity of patients makes it difficult to design efficient clinical trials. A score like TREWScore could help identify cohorts of patients for inclusion into clinical trials. Moreover, by identifying patients with reasonable specificity many hours prior to the onset of septic shock, TREWScore could enable clinicians to investigate the benefits of proactively starting treatment prior to shock onset. This has yet to be studied, in part due to the lack of a highly sensitive and specific tool to identify at-risk patients.
Along with my collaborators, I am in the process of launching a pilot study to prospectively validate TREWScore. In the future I hope to expand the methodology developed here to other diseases. In addition to disease prediction, I research how to estimate the effects of treatments on patients with the goal of providing better decision support tools to clinicians. I benefit from close collaboration with clinicians at Hopkins and it is important to me that my research is clinically meaningful.
I am currently a PhD student in computer science at Johns Hopkins University and an NSF graduate research fellow. I received my bachelors in computer science and linguistics from the University of Chicago and my early research focused on low-resource speech recognition. However, when I started my PhD in computer science at Johns Hopkins University, I was introduced to medical data and quickly became fascinated by how computer science could potentially be used to improve healthcare and provide clinicians with better tools for patient care.