“Clinician burnout is a critical issue to understand and address,” says Mohsen Bayati, a professor of operations, information, and technology at Stanford Graduate School of Business. The condition is thought to affect nearly half of all U.S. doctors, at a cost of about $4.6 billion annually due to turnover and reduced work.

Doctors report that one of the major reasons they feel burned out is electronic medical records, the digital records that have replaced paper patient files. It’s not uncommon for doctors to spend more time updating EMRs than spending time with their patients.

Yet EMRs also contain valuable information that could predict physician burnout as it’s happening. In a recent study, Bayati and his collaborators Daniel Tawfik, Jessica Liu, Tait Shanafelt, and Jochen Profit of Stanford University School of Medicine; Amrita Sinha of Harvard Medical School; and Thomas Kannampallil of Washington University School of Medicine, used AI to analyze routinely generated information from EMRs to predict the likelihood of burnout, particularly at the clinic level. This innovative use of medical records enabled the researchers to link concurrent predictive measures to burnout rather than looking for possible connections after the fact.

“The current way we identify and respond to burnout is always retrospective,” says Tawfik, an assistant professor of pediatrics. “It’s done through surveys, and by the time those are administered, respondents are answering based on things that happened in the past. We wanted to find a way to prospectively identify conditions that caused high risk for healthcare worker burnout, to act before the problems become apparent.”

“Surveys are useful for telling us who has burnout,” Bayati says. “The medical record can’t tell you that, but we can use it to predict the clinics most likely to yield burnout and take steps to address that.”

Inbox burden and other symptoms

The researchers used machine learning to identify links between elements of EMRs and burnout in a sample of 233 physicians from 60 clinics over 18 months. The initial data set contained over 1,500 potential predictors of burnout, which the team winnowed down to about 200.

“We used clinicians’ expertise to audit the variables for those more relevant to burnout, creating a more powerful combination of human judgment and automated algorithm,” Bayati explains. “Typical research might just throw all the data into an algorithm, which might result in overfitting” – drawing connections from the data that may not be valid.

The variables the researchers ultimately examined included doctors’ clinical workload, such as their number of appointments and the complexity of their patients’ cases, and measures of efficiency, such as how long clinicians took to place an order for medical tests or to write a note in a patient’s file. Another variable was how long doctors spent reviewing charts, Tawfik explains. “How many minutes was the provider spending actually looking through the patient’s prior records before their appointment?” The researchers also measured the sentiment of doctors’ notes, as reflected in their wording and tone.

The collaborators then applied multiple algorithms based on different machine learning models to link these variables to a “burnout score” based on surveys filled out by the doctors. “These are surveys administered every 12 to 18 months across the organization to check well-being as part of quality improvement initiatives at the clinics,” Tawfik says.

The study revealed which measures best predicted burnout at the individual and clinic levels. The most predictive variable was the number of automated messages a physician receives. “All sorts of messages come into the clinician’s inbox in the electronic record, similar to email,” Tawfik says. “Things they need to respond to and act upon, creating a sort of inbox burden.”

The next-best predictor of burnout was counterintuitive: when clinicians have team members write some or all of their medical notes for particular cases. “Our interviews with physicians showed that having other people write the note doesn’t actually speed up the workload because they find that they have to go back and edit the note,” Tawfik says. “Many feel they can write the note faster themselves.”

Prescriptions for treating burnout

Overall, the EMR-based predictors of burnout were most robust at the clinic level. “We found better performance of the algorithms when it came to predicting where a particular clinic would rank among all organizations from the highest burnout to the lowest burnout levels,” Tawfik says. There was what he calls “only a limited signal” in the EMRs for predicting individual clinicians’ risk for burnout: “We found a somewhat muted but present ability to predict individual burnout scores.”

The results point to a potential method to identify clinics where doctors are at risk of burnout without relying solely on survey responses or other inputs from the physicians. “You can use evidence from the electronic records to go to higher-risk clinics, identify the largest pain points, and try to generate interventions to prevent burnout,” Tawfik says. Bayati adds, “All research of this type would benefit from a more real-time measure of burnout, rather than surveys alone.”

Specific interventions could include improving the signal-to-noise ratio in clinicians’ inboxes. “Our conversations show there are some messages physicians find highly valuable and appreciate receiving, and others they don’t find valuable at all,” Tawfik says. “But a message one physician finds valuable may be an item another doesn’t value.” Enabling individual clinicians to customize which messages get through to them could reduce the stress of inbox management.

Likewise, some physicians prefer team-based notes while others don’t. “It seems there’s a learning curve for this,” Tawfik says. “There’s an initial period where it may take a clinician longer to write the note as part of a team, but if you stick with it and you have people who really learn how to write the notes for you, then the team-based approach can be beneficial.” Again, an individualized approach where physicians find the note-taking strategy that works best for them could reduce their chances of burnout.

Looking ahead, Tawfik says the next step is to develop more accurate ways of screening doctors for imminent burnout – without adding to their workloads. “Real-time risk indicators could help us identify high-risk settings with less survey burden for clinicians,” he says.