The Problem
Reducing in-hospital mortality is a key quality and safety priority across hospitals in the United States. Unfortunately, an average of 2% of patients admitted to US hospitals die during the inpatient admission. For some patients, particularly those with terminal illnesses, dying is something that is expected and planned for over a period of weeks to months, or even years. For other patients, particularly those with acute illnesses, death can be prevented with hospitalization and aggressive treatment. To date, efforts to reduce preventable in-hospital mortality have focused on improving treatments and care delivery, and efforts to reduce non-preventable mortality have focused on supporting patient preferences to die at home and attempting to reduce health care costs in the inpatient setting. Early identification of patients at high risk of inhospital mortality may improve clinical and operational decision-making and improve outcomes for these patients.
Our Solution
To assist efforts to improve the quality and safety of care at Duke Health, we partnered with the DUHS Mortality Review Team to build a machine learning model to predict in-hospital mortality, run at the time of admission to the hospital for all adult patients. We carefully designed the model to be implementable on a system-level, choosing an approach that was not disease specific, used accessible computational methods, and relied on data readily available in EHRs. We retrospectively evaluated model performance at DUH, DRH, and DRAH, and completed the evaluation of a machine-learning model to predict in-hospital mortality with highly encouraging results that we plan to publish. We created a model facts sheet, similar to a drug label, that clearly explains the “indications” and “contraindications” for model use and provides other important information for clinical and operational leaders. Lastly, we prototyped initial workflows to test, established baseline metrics, and built a dashboard to display patient risk scores to support initial workflows. We plan to evaluate different workflows across DUH, DRH, and DRAH to assess the impact on patient outcomes, and look forward to sharing the results.
