Applying a Human Touch to EMRs and AI

Both electronic medical records (EMRs) and artificial intelligence (AI) are key elements of healthcare’s technological revolution, and their advantages and drawbacks often intertwine. AI relies on data often contained within the EMRs, and it may be able to help to assuage some of the difficulties that EMR use has presented to physicians.

In a new commentary written for JAMA, Stanford University’s Abraham Verghese, MD, and colleagues assert that the 2 technologies can be made harmonious, but only if proactively guided by human physicians.

The adoption of EMRs has been rapid, thanks to subsidy and policy, but the efficiencies that the systems are meant to create have not materialized for many physicians. As Verghese and his coauthors note, their use has created additional data entry work for countless physicians, at the cost of human interaction between physicians and their colleagues and patients. The “4000 keystrokes a day” problem has also “contributed to, and perhaps even accelerated, physician reports of symptoms of burnout,” they write.

These issues dovetail with AI’s growing capabilities and acceptance: Predictive machine learning could help physicians make better, more accurate prognoses and steer treatment towards more positive outcomes, but the physicians have to be feeding good data into the EMR for that to happen.

Algorithms will replicate human biases—and even exaggerate them—if those biases exist in the source data. Information entered without context can also lead to insights that aren’t very useful: The authors use an example of a machine predicting better pneumonia outcomes in patients with comorbid asthma, because the data it was given lacked the context that the asthma was the very reason physicians admitted a patient early and gave them particular care.

“The missing piece in the dialectic around [AI] in health care is understanding the key step of separating prediction from action and recommendation,” the commentary argues. Models that can't explain causation or their underlying processes are not useless: That idea should be abandoned. Instead, the authors say, medicine should allow the machines to predict while leaving the interpretation and decision in the hands a physicians.

Physicians and computers have to work together, they write, to compensate for the other’s deficiencies. The algorithms can’t steer themselves, but they can provide insights quicker and more accurately than any human brain.

Machine learning technologies like natural language processing (NLP) could also make the EMR less of a drain on physicians, by both entering in many of the required notes and extracting data that could be useful for analytics. Automated processes could restore some of the time and humanity that many physicians say they sacrifice while filling out EMRs.

Verghese and one of his co-authors, Nigam H. Shah, PhD, also collaborated on a previous JAMA commentary in 2016 about the burdens presented by EMRs. Both that article and the new piece, published this week, reference Frances Peabody’s famous “The Care of the Patient” lecture: “the secret of the care of the patient is in caring for the patient.” It will be much easier for physicians to do that, they write, if human intelligence and AI can collaborate to produce “a well-informed, empathetic clinician armed with good predictive tools and unburdened from clerical drudgery.”