AI v doctors: If ChatGPT can diagnose patients, why do we need doctors?

If AI models can diagnose patients in seconds and large language tools such as ChatGPT are offering patients diagnoses at their fingertips, where does that leave doctors?

The rapid evolution of AI has put medicos on notice: the future is here and they had better get on board or get left behind.

Almost every week, it seems, there are stories of newly developed AI models besting doctors. And while there’s little doubt the process of determining a medical diagnosis is already changing, don’t count the humans out of the loop just yet.

AI might come out on top when harvesting and analysing swathes of data but there’s still a long way to go — it’s doubtful the robots are ever going to put doctors out of a job.

Is AI better than my doctor?

An AI system developed by the boffins at Harvard University dubbed Dr CaBot has shown what is possible.

In July, when was pitted head-to-head with an expert diagnostician in a demonstration that was likened to the 1997 chess match between chess Grand Master Garry Kasparov and IBM supercomputer Big Blue (which won), Dr CaBot was astonishingly quick and accurate.

Dr Daniel Restrepo was given six weeks to prepare his diagnosis presentation. Dr CaBot got six minutes.

Given a case study of a 41-year-old man who had come to the hospital with 10 days of fevers, body aches, swollen ankles, a patient rash and two instances of fainting, Dr Restrepo diagnosed the patient with Lofgren syndrome.

Then it was Dr CaBot’s turn. In a warm woman’s voice, it talked through its diagnostic reasoning, generating a range of possibilities and noting their strengths and weaknesses.

“Putting it together, the single best fit is acute sarcoidosis manifesting as Lofgren syndrome,” it concluded.

In China, they are having equally incredible results.

At a Shanghai medical AI skills competition on November 20, four masked physicians from top Shanghai hospitals split into two teams to face-off against two different AI systems — a gastrointestinal multimodal AI and an international AI system — to diagnose a real-life case.

The AI produced diagnoses in under two seconds while the doctors took 13 minutes, China’s state-run media reported.

The international model reportedly fell slightly short on diagnostic accuracy, while the Shanghai-built gastrointestinal AI model, which was trained on 30,000 real cases and can read endoscopy and CT scans, matched the diagnosis of the doctors and their recommendations for treatment.

So, what does that all mean for the future of medical diagnosis?

Well, AI won’t be replacing doctors anytime soon. But those doctors who embrace it will definitely replace those who don’t and they will reap the benefits of improved diagnostic accuracy, reduced administrative load and better triaging of patients.

Why shouldn’t I use ChatGPT to diagnose myself?

It might be slightly better than diagnosing yourself from a TikTok video or an Instagram reel, but Open AI’s ChatGPT and large language models like it are not fool proof and in fact can get things just plain wrong.

While people have been consulting “Dr Google” for years, these tools bring something different to the table. A feeling that for some mimics a doctor’s consultation.

Overconfidence of some of these AI tools can mean that they take patients down the wrong path and down rabbit holes

A patient lays out their symptoms and the chatbot gives its best answer, specific to what it is asked — not just a list of documents that might be helpful as with Google — to diagnose what ails them.

“It’s nothing new for us really in general practice that patients have sought their own information out before they come to see us . . . but obviously AI has changed the tools that patients are using and how they access information,” GP and Australian Digital Health Agency chief clinical advisor Dr Amandeep Hansra told The Nightly.

“They might go to ChatGPT or other large language model AI tools and try and work out what they think the doctor’s going to say or what they might suggest as a diagnosis.

“The core relationship hasn’t changed. I still need to help them interpret the information that they’ve brought to my consult, and I still need to continue contextualise and personalise that care.”

Depending on the information you feed into it: age, gender, past medical history. The answers you get back can feel like it’s talking specifically about you, meaning patients are more likely to believe it is 100 per cent right.

But that’s not always the case.

“Overconfidence of some of these AI tools can mean that they take patients down the wrong path and down rabbit holes,” Dr Hansra said.

She recalled an experience with one patient who was convinced the diagnosis and recommended treatment that she gave for a common ailment was wrong because ChatGPT had told them otherwise.

“I met the patient did my normal consult and the patient had already looked to ChatGPT and decided what treatment they wanted. They wanted a specific medication which I didn’t prescribe for them.”

Dr Hansra had to explain why she was right and ChatGPT had it wrong.

The difference, she said, was that she had examined the patient and the AI tool hadn’t.

“Patients forget that when we’re assessing them, we’re taking a lot of different things into consideration and also our experience. I’ve been a GP for 20 years, that experience counts for something versus what an AI tool can find on the internet.”

“I think it’s really important that people do check anything that they’ve got from these AI tools with their regular treating practitioner because I have seen these tools be super confident at things that are very clearly wrong.”

Even Open AI founder Sam Altman agrees. In August he told a conference in the US that ChatGPT was already surpassing human doctors in diagnostic accuracy.

“ChatGPT today, by the way, most of the time, can give you better – it’s like, a better diagnostician than most doctors in the world,” he said.

“Yet people still go to doctors, and I am not, like, maybe I’m a dinosaur here, but I really do not want to, like, entrust my medical fate to ChatGPT with no human doctor in the loop.”

Australian e-Health Research Centre director Dr David Hansen said understanding exactly how large language models such as Chat GPT arrived at their answers was really important.

“It’s been trained on a lot of data and it will use a statistical method to form sentences based on the input, and that seems to work really well most of the time,” Dr Hansen told The Nightly.

“But also don’t take it as gospel. It’s just as important to make sure you do go through the normal process of talking to a real doctor, who may also be using AI tools but they’re doing it in in educated way.”

Dr Hansen said the Health Department’s Healthdirect symptom checker was probably a better tool. He said that had been trained on medical data and was continually revised and checked through rigorous testing.

How is AI listening in on my consultation?

Beyond diagnosis, for some doctors, other tools such as ambient medical scribes are helping keep them in the profession.

Probably one of the most noticeable incursions of AI into a doctor’s room, these tools free doctors from the need to take manual notes and allow them to look their patients in the eye during a consultation.

Developmental paediatrician Kim Drever was on the brink of bowing out in mid-2022 because the COVID pandemic and a growing roster of patients who were struggling had piled on the pressure.

Her job requires her to take extensive clinical notes and consult broadly with other professions, including GPs, schools and the justice system and she’d become burnt out under an avalanche of paperwork and a system that was jammed.

“I just felt I couldn’t be helpful anymore. . .(my job) is to help them get care in their community, the mental health or, you know, extra school help, or whatever it might be.

“I could not keep up and I just thought I can’t be helpful, I’m done, this is awful.”

But an AI scribe tool helped keep her in practice.

The medical scribe “gave me back my weekends and it also meant that I felt like I could be helpful because I could generate useful letters quickly and it could get things started for that family quickly”.

Up until recently, Dr Drever was an advisor to Lyrebird Health, one of about half a dozen AI scribe companies in Australia, working with the engineering team to help tailor it to the needs of those in her specialty, outlining the rules and scope of how it should work.

“From the very get-go, the potential of it was amazing, but it is no good to anyone if the design of it and the intent of it is not informed by people who are using it,” she told The Nightly.

“Scribes or anything that is kind of partnering with a doctor needs to act like us, but not think for us.”

She also noted it was important patients be made aware when AI was being used and to what degree doctors were relying on it.

Why is it risky diagnosing yourself with ChatGPT?

What the humans need to do, Dr Drever said, was to bring context to the huge amounts of data being taken in by AI tools.

There are lots of different ways AI can be inaccurate, including hallucinations, where it completely makes something up. AI omissions, too, can be just as damaging.

Dr Drever gave an example where she was using an ambient scribe during a consultation when a child made disclosures about being sexually abused.

Relying on the scribe to take accurate notes that may one day be used in court proceedings, Dr Drever was alarmed to find it wasn’t taking down the sensitive remarks because of restrictions imposed by Open AI and Amazon.

This has since been resolved but could have been disastrous if Dr Drever hadn’t resorted to transcribing by hand.

“For me that omission could have been harmful to that child, right? Because then you’d have to rely on my memory and maybe I didn’t remember properly because you’re relying on another tool that failed you.”

There are also inaccuracies where the AI tool can “hear” things wrong, such as a blood pressure reading or a person’s name.

And if the AI gets it wrong and that has an impact for the patient, then it’s the doctor that could be held legally liable.

What is AI’s potential for good?

That aside, Dr Hansen is excited for what the future holds when it comes to AI in medicine, especially around genomics, imaging and precision medicine — finding the right treatments for individuals for everything from cancer to rare diseases.

“That’s an area where clinicians. . . really need to be leveraging what AI can offer to make sure that we’re getting the best diagnosis as fast as we can.”

Dr Hansen said he expects AI tools will also be doing more triaging of patients.

“You could do an AI consult beforehand and they could give you the blood test order and you go get your blood test before you go to the doctor so you can discuss the result,” he said.

“So there’s lots of efficiencies that we can think of like triaging people and making sure they’re going to the right place.”

But Dr Hansra is sceptical that AI will ever make doctors redundant, because to do so would be high risk.

“AI is really great at automating tasks and helping us sift through large amounts of information but I think health care itself relies on clinical judgment, ethical decision-making, empathy, understanding of the complexity of people’s lives. . .all of those factors that come into our decision-making.

“I think the AI will help augment what we do, but I don’t think it can replace what we do.”