Healthcare will see large productivity gains from artificial intelligence, but not through robotic medical services right away. The big, early gains will come from faster, easier clerical tasks that today consume too much of doctors’ and nurses’ time.
Healthcare services constitute 11% of gross domestic product. That figure rises to 18% when non-service spending is included, such as drugs, devices and construction. A big gain in this industry will really help total output of the economy. Healthcare staffing shortages accentuate the benefits that will accrue from productivity gains.
I previously expected slow take-up of AI tools in healthcare, but I’ve changed my opinion. Although it’s still true that healthcare has very high stakes, high litigiousness and a good deal of regulation, there are many clerical and administrative tasks that can be automated fairly easily.
AI for Clerical Healthcare Tasks
Eren Bali, the founder of Udemy and now founder/CEO of Carbon Health, explained in a conversation the large opportunity, using his company’s charting software as an example. Doctors and other direct providers spend a great deal of time on clerical tasks, such as writing up patient visit summaries. At Carbon Health, doctors have the option of using an app on their phone or iPad that will record the conversation. The recording is transcribed, then GPT-4 creates a summary of the patient’s concerns, doctor’s observations and treatment plan. The doctor edits the health record as needed, then signs off on the report.
Electronic health records offer substantial potential benefits, but practitioners have complained about the time required to write up patient visits. Carbon Health reports time to fill out the patient chart has dropped from 16 minutes to less than four minutes, with much greater detail. Ponder that gain. In terms of the physician’s time, that’s a huge savings. In terms of the physician’s job satisfaction, less paperwork and more patient time is a big plus.
Up and down the healthcare system, doctors, nurses and clerical personnel perform simple tasks that do not involve diagnosis or treatment of illness. But the tasks are often vital for good patient treatment or for billing. Using AI makes tremendous sense. Carbon Health’s next steps include easier writing of prescriptions, lab orders and patient communications.
Carbon Health is not alone in this market. Other firms have developed scribing tools, as they are called, that incorporate artificial intelligence for transcription and summaries.
AI in Medical Diagnosis and Treatment
The use of AI in diagnosis and treatment is far more controversial. The Wall Street Journal reported on dissatisfaction among nurses about AI systems conflicting with the their views of what was best for patients. Some of the criticisms by nurses reflected poor management decisions more than bad AI. Nurses have been told to both use their judgment and to listen to the AI’s recommendations. Every company with a procedures manual has had the same problem: When should an employee go by the book, and when should the employee use judgment because the book isn’t right in a particular circumstance?
Right now ChatGPT seems to be about as good as a newly graduated M.D. The AI got a 60% score—a passing grade—on a simulation of the United States Medical Licensing Exam. (We patients may be worried that our doctors could pass the exam with 40% incorrect answers. However, the exam is typically completed in the doctor’s first year out of medical school, before completing a residency. So the physician we see in a clinic usually has more experience.)
The likely next step in AI usage for diagnosis and treatment is a “recommendation engine.” The AI would listen in on the conversation between patient and doctor while accessing past visits, lab reports and imaging, and occasionally make a recommendation. “Recommend you consider possibility of infection as a cause of patient’s vision loss” or “Recommend an x-ray to rule out possibility of a fracture.” The practitioner could reject the recommendation or accept it. This process would combine the knowledge of a trained physician or nurse with the AI, giving the human final authority. The recommendation engine could be trained on actual patient data about practitioners’ final decisions and the patients’ outcomes.
Over time, recommendation engines will improve, possibly to the point where the recommendation engine can be used where practitioners are not readily available. That could include rural clinics or after hours care. We may find that a good nurse with a good recommendation engine can substitute for a doctor in many cases. However, the transition will be managed cautiously given that human lives are at stake.
Healthcare Productivity Gains from AI
Medical care includes hands-on activity, which will be slow to benefit from AI. However, much healthcare involves processing information about a patient and the medical possibilities. Over 1.1 million people work as physicians, surgeons, physician assistants and nurse practitioners. If each of these can save a few minutes every hour through greater productivity, the country will reap large gains.