A staggering 4.6 billion dollars are spent annually replenishing the ranks of physicians who depart from their current positions. High physician turnover is detrimental to the quality of patient care and contributes to the notorious inefficiency of the healthcare system. However, in a paper recently published in the journal PLOS ONE, researchers at the Yale School of Medicine have developed a model to identify when a physician may be at risk for departure. “As a clinician myself, […] what brought you into medicine is the time you spend talking with patients and hearing their stories, but often what takes up a lot of your time is documenting, or finding the correct order that’s going to be covered by insurance,” said co-author Andrew Loza, who earned his Ph.D. from Washington University in 2016 and later obtained his M.D. from Yale School of Medicine.
The study aimed to develop a model to predict physicians who may be at risk of departure by identifying and quantifying factors that interfere with a physician’s ability to work directly with patients. The researchers used an algorithm called XGBoost, which receives a set of input variables to classify the level of a physician’s departure risk. Some of these variables included electronic health record (EHR) use, tenure (time since hiring date), physician age, and patient volume. However, due to the vast number of parameters in the model, it is difficult to understand the recipe that turns input variables into an output. So, the researchers used a data analysis technique called Shapley Additive Explanations (SHAP) to determine the level of influence of individual variables on a physician’s departure risk, and subsequently derive conclusions from the data.
Interestingly, tenure was found to be the strongest predictive factor in classifying physicians at high risk of departure. For physicians with longer tenures (10-35 years), high EHR use was found to increase risk of departure and decrease risk for less experienced physicians. Similarly, longer documentation times were shown to reduce the risk of departure in some tenure brackets and elevate the risk in others. Since this was an observational study, the algorithm cannot establish a causal link between the identified factors and physician departures. Loza provided an example illustrating the fallacy that correlation implies causation. “If you look at literature for retention of software use, individuals who report software bugs are more likely to keep using the software, but you would never tell a manager I think we should add more bugs,” said Loza. Further work is currently being conducted to uncover any potential underlying causal relationship between the identified variables and departure risk.
Ultimately, Loza hopes that this research will serve as the foundation for developing screening tools to identify physicians who are at risk of departure and provide support for them to better engage with patients. “I hope this would be used to shift the variables which are associated with risk of departure into lower risk and increase the time spent with patients,” said Loza.