Predictive models can augment a clinician’s bedside assessment to help identify patients who are deteriorating. We collaborate with clinicians, researchers, and healthcare organizations to ensure that predictive models are tuned to each patient population. We’d like to provide perspective from two Epic clinicians on how these tools came to be and how they can improve clinical care.
We’ll start with a discussion of sepsis. It’s an infection that results in organ failure and, often, death. Clinicians reading this might remember when many of us were using the Systemic Inflammatory Response Syndrome Criteria (SIRS+) as a way to identify and screen patients who might have sepsis. Many recognized that SIRS was not an ideal tool, but it did help identify some patients with sepsis sooner than a clinician’s eye.
Patients can meet SIRS criteria for many different reasons, including labor or pregnancy. It was too broad of a tool and left many feeling frustrated. To help improve on this, Epic used robust patient datasets and predictive modeling techniques to perform a similar assessment to SIRS using different criteria. We created a predictive model that identified more patients correctly and reduced the number of times an alert was triggered when compared to SIRS. The predictive model both improved the value and decreased the noise for clinicians when identifying patients with sepsis.
Predictive models require analysis and tuning to work properly in your clinical context. We created a tool called the validation utility for every health system to see exactly how effective Epic’s Early Detection of Sepsis model was at identifying septic patients in its population.
Since then, we’ve scaled that process to a number of different models, and many organizations have found them very helpful. Some have shared their success at our XGM and UGM conferences, and some have published in the medical literature. Organizations like Prisma and North Oaks have seen improvements in sepsis mortality using these tools. Tens of thousands of clinicians have access to the sepsis model and transparency into how it works. They can hover to see what parameters it takes into account, and the details of the model’s development are available to Epic organizations. However, it’s not just about the inputs and the math. The robust clinical workflows and processes that surround these tools are what give the tools purpose and allow for improved outcomes.
We recognize machine learning is not a perfect solution for every problem, but we believe it has a real opportunity to make healthcare better. As we continue to improve machine learning at Epic, we will further build on our collaboration with the clinical and research community and are always looking for ways to partner with clinicians to create models that can improve clinical care and promote better health for everyone. If you would like to engage with us, reach out on the UserWeb.
By Jackie Gerhart, MD, & Johnston Thayer, RN – Clinical Informaticists at Epic
