Yale model identifies risks that can lead to physician turnover

A machine learning model used by Yale University researchers has identified various elements associated with a higher risk of physician turnover, a disruptive and costly problem in the healthcare industry.

YaleNews reports a physician's length of tenure, age, and the complexity of their cases were among factors causing them to leave their job.

“There have been efforts to make machine learning models not black boxes wherein you get a prediction but it’s not clear how the model came to it,” said Yale School of Medicine clinical informatics fellow and study co-senior author Andrew Loza, according to YaleNews. “Understanding why the model produced the prediction it did is particularly useful in this case as those details are going to identify issues that may be leading to physician departure.”

YaleNews states the researchers were able to predict the likelihood of physician departure, with a 79% accuracy rate, by analyzing data from a large U.S. health care system over a three-year period.

Although traditional surveys are often used to track physician burnout and job satisfaction, they can have low response rates and may not provide insight into the long-term behavioral patterns of physicians. 

The Yale researchers addressed these limitations by using data from electronic health records (EHRs) to observe physician behavior moment-to-moment over a long period of time. The EHR data included clinical productivity measures, physician characteristics and how much time physicians spent using EHRs.

The researchers used the EHR data to train a machine learning model that could predict whether a physician would leave their position within the next six months.

“The findings highlight there's not a one-size-fits-all solution,” Loza said, as reported by YaleNews.

The researchers created a dashboard that can display this information and help healthcare leaders intervene before physicians decide to leave, ultimately reducing physician turnover.

The study was published in PLOS ONE on Feb. 1.