University of Minnesota researchers have developed a machine learning algorithm using electronic health record data to improve care delivery for seriously ill patients by accurately predicting the risk of 1-year mortality.

The random forest (RF) model, which estimates the risk of death within a year of the last day of hospitalization, leverages commonly obtained EHR data such as vital signs, complete blood count, basic and complete metabolic panel, demographic information, as well as ICD codes.

Nishant Sahni, MD, adjunct assistant professor in the Department of Medicine at the University of Minnesota Medical School, sees the model as a potential clinical decision support tool for improving end-of-life planning.

“Having been a hospitalist for more than 15 years, I find that the end-of-life conversations don’t necessarily happen between clinicians and patients,” contends Sahni, who notes that seriously ill hospitalized patients are frequently subjected to unnecessary, invasive procedures that do not enhance their quality of life.

Nishant Sahni, MD
Nishant Sahni, MD

“Unfortunately, we don’t have a lot of prognostic models that just look at a general hospitalized population,” adds Sahni. However, with the machine learning algorithm, he says “the hope is when a patient is leaving the hospital, the physician will get a notification that the patient is high risk and needs those specific conversations and care, which would empower patients to make more informed decisions regarding their medical care."

Results of a retrospective study, using EHR data collected from almost 60,000 hospitalizations linked to the state death registry, were published last month in the Journal of General Internal Medicine. “Age, blood urea nitrogen, platelet count, hemoglobin and creatinine were the most important variables in the RF model,” state the study’s authors.

Also See: Data analysis shows ICU scores accurate in predicting risk of death

“We’re trying to find partners at this point,” according to Sahni, who says the goal is to figure out how to deploy the model in an EHR so that clinicians are not required to enter any data manually. In addition, he adds that a web-based application programming interface (API) design is available to enable developers to query the model from anywhere and use it in their development of apps.

The University of Minnesota is interested in licensing the technology for further research and development. “The license is available for this technology and would be for the sale, manufacture or use of products claimed by the issued patents,” states a web posting. Additional information on the licensing opportunity for commercial purposes is available here.