The AI physician will see you now

or in 2035

Photo by Emilipothèse on Unsplash

In October, 2024, I sat in a large cool Hyatt conference hall in San Francisco with hundreds of physicians and some physician associates and nurse practitioners. We were about to watch a heavyweight matchup. In one corner stood a suited, six-foot bespectacled Indian man, and America’s most famous diagnostician, Gurpreet Dhaliwal. In the other corner, a faceless entity who was nowhere but everywhere: Chat GPT-4.

Two large power point slides hung in front of the black backdrop. Two people stood on stage. Bob Wachter, the chair of the UCSF internal medicine department, introduced Dhaliwal. A physician and professor of medicine who works in the ER and hospital ward, Dhaliwal gets on stage every year at the annual UCSF hospitalist conference and Wachter presents an unbeknownst case to him.¹

Working through a case in front of someone, let alone hundreds of healthcare providers, is a physician’s nightmare. Why? Because we all have limitations in our ability to process information and diagnose. We have blind spots. We have biases. But Dhaliwal is willing to get up on stage and show his thinking aloud.

Wachter presented the case.

“This is a 27 year-old woman with recurrent pancreatitis who presents with a two-day history of severe abdominal pain, nausea, vomiting, and diarrhea.”

Dhaliwal explains his initial thinking based on her age and symptoms. He discusses how the word “recurrent” could be misleading or false and maybe her pancreatitis never went away in the past.

“This could be as simple as pancreatitis happening again, but it could range to something like gastroenteritis or a different infection like appendicitis.”

Wachter revealed more about the woman’s medical history. She had three prior episodes of pancreatitis in the past eight months. She stopped drinking alcohol after the first episode. She’s now lost 50 pounds over six months.

Dhaliwal expanded his diagnoses to include autoimmune pancreatitis.

Labs are shown from her previous episodes of pancreatitis. Imaging showed no gallstones. On exam, the woman has tenderness in her abdomen. Imaging now shows innumerable liver and bone lesions and nodules in both her breasts. The pancreatic duct is now obstructed.

Dhaliwal explains his thinking.

“The pancreas is sort of ground zero for her illness. I’m thinking about lymphoma from her pancreas that may have spread. The other way to think about this is that the disease started in a different organ and it may have spread to the pancreas, though that’s a little rarer since the pancreas is rarely the recipient of other diseases. But I suspect it started in the pancreas. When I think of cryptic cases that started in the pancreas, I think of genetic disease like cystic fibrosis, or autoimmune disease. I’m thinking of sarcoidosis. Lymphoma would be strange because there are no enlarged lymph nodes but it’s possible. IgG4 pancreatitis is possible but IgG4 rarely takes over the liver and usually has no predilection for the bones, so I don’t think it’s that.”

As I sat in my chair, arms folded and laptop closed, I marveled at Dhaliwal’s mind. Nowhere in my brain did I keep data like Dhaliwal. His diagnoses list was expansive yet practical. The fact that he knew the tendencies of IgG4 pancreatitis struck me. His statement also of the pancreas rarely being the receiver of other disease showed me how great his pattern recognition was. Dhaliwal had a mind palace and behind each door was a set of characteristics he knew about a disease.

After some back and forth, Dhaliwal settled on a diagnosis.

“In this case, a biopsy will give the diagnosis. The fact that her disease exploded on her imaging to me suggests lymphoma over sarcoidosis or IgG4-related pancreatic disease.”

Before the case started, Dhaliwal emphasized that the point of the case isn’t to get the correct answer, but it’s to show the intricacies of diagnosing.

Wachter then pulled up Chat-GPT4.

“How are you?” he asks.

“I’m good,” it replies on the loudspeaker.

“I’m a doctor. We’re here with 600 other people discussing a case. I’m hoping to get your help regarding a 27 year-old woman.”

Wachter explained the scenario. GPT-4 provided considerations.

“Given this young woman’s complex case, we should consider a diagnosis that ties together her recurrent pancreatitis and her hepatocellular weight loss.”

GPT-4 considered multiple diagnoses and gave their differential diagnosis: IgG4 pancreatitis, autoimmune pancreatitis, and malignant pancreatitis. As it mentions malignant pancreatitis (pancreatitis from cancer), it includes pancreatic adenocarcinoma (most common) and neuroendocrine tumor (less common) as possibilities.

Wachter asked everyone if they were ready for the answer. The next slide revealed the diagnosis.

“The biopsy of the pancreas shows a neuroendocrine tumor,” he said.

The audience collectively gasped.

“The patient is seeing a neuroendocrine GI tumor specialist.”

In Dhaliwal’s defense, he was bold enough to choose a diagnosis. GPT-4 listed several and managed to capture the right diagnosis. GPT-4 sucker punched Dhaliwal when he wasn’t looking. Cowardly, but smart move.

---

So what does this mean? Are doctors destined to be replaced by AI?

It’s a sexy question because it’s threatening and futuristic. Based on Dr. Wachter’s next talk, I came away with an answer.

Digitalization in medicine happened only recently. The paper chart was replaced by the electronic medical record (EMR) in the early 2000’s. I started medical school in 2010 and trained with the EMR but I overheard groans from older physicians who didn’t want things to change. They were already managing a lot; learning a new way to document meant more work.

Wachter showed a picture of a drawing by a child, who recalled their doctor’s visit. In the drawing, the physician is sitting, facing the computer, away from their patient.

“Physicians became expensive data entry clerks,” he said. “They became scribes to better bill patient care.”

With the introduction of EMR, the inbasket was created, where labs, test results, and patient messages ended up. After clinic, physicians would work after hours to finish notes, address lab results, and answer patient questions.

“The EMR inbox became inundated with messages. None of us anticipated this, but it’s logical that this is how it would happen, when you give patients labs they don’t know how to interpret, and now you can message your doctor for free. It led to 24/7/365 access without thinking about the business model.”

I suspect somebody knew this would happen. Perhaps the executives who look at physicians and healthcare providers as productive puppets, wanting to increase the eternal churn of seeing patient after patient.

Wachter then showed how AI evolved from the early 1970s to now generative AI. Generative AI has gotten smarter. It has setbacks like hallucinations - producing non-sensical or incorrect information - but these mistakes are being fixed in a matter of months. He showed how AI can be used in medicine, to look up literature in rapid pace, and how it can do mundane tasks like creating physician notes and medication approvals.

He ended the talk with quoting the dean of USC’s medical school: “AI won’t replace the physician. But physicians using AI will replace physicians not using AI.”

---

I walked away from that talk relieved. I’m not going to be replaced by AI anytime soon. It can do party tricks like arrive to diagnoses quickly, but it can’t do everything physicians do. I don’t know the trajectory of AI in medicine in the next decade but I’m safe for now. My time in San Francisco gave me permission to use AI in my work in Madison. If experts like Dhaliwal are using it, I’m going to use it.

I do wonder though what the powers that be will do with AI. With digitalization in the form of EMR, productivity can be more easily measured. With AI, I wonder if productivity will be expected to be increased, as profit is the main motive across so many fields, including media.

---

How I Write with David Perell has become my favorite writing podcast. His guests are writers and he asks them great questions about how they generate ideas and work on their craft. He recently had TV show writer

Michael Jamin

on and during the end of their talk, they touched on AI in writing. Perell, playing devil’s advocate, was for writing with AI. Jamin didn’t see the need for it.

“AI is not a tool, it’s in lieu of. I don’t get in a Waymo and say ‘I’m a great driver.’ The car is a great driver. Writing is hard, it’s supposed to be hard. If you’re using AI to be a writer, why did you become a writer?”

I was on Team Jamin. Then, he made a great point about art.

“None of us enjoy writing. All of us enjoy having written. The point of art is to take something ugly or painful and transform it like the way you make mud into a vase. That’s the point of it.”

He shifted to what studios - production companies - would do if they had their way.

“Studios would love to replace writers with AI,” he said. “That way they can fire all the writers. They’re always looking for ways to fire writers.”

The big studios he’s referring to are ones like Paramount or Universal, massive companies that would love to help their bottom line by using AI. I don’t like this idea as it means many people being displaced from their jobs while people at the top continue to do well.

After listening to Jamin, I realized that it’s not AI that I fear, but what one powerful person, usually male, will do with AI - replace jobs, decrease a person’s ability to afford their rent and groceries, reduce job quality. If the current political administration is any clue as to what one person with excessive power can do, then the misuse of AI by man is the truly scary thought.

---

1. Journalist Katie Hafner (who is married to Wachter) wrote about Dhaliwal in a New York Times article in 2012.

---

Thank you to Gurpreet Dhaliwal, who generously reviewed this essay to ensure its accuracy. His response to reading the essay: “I'm a fan of AI, use it everyday, and have written a number of supportive commentaries, and yet agree with your assessment: It can do party tricks like arrive to diagnoses quickly, but it can’t do everything physicians do."