The next challenge for ChatGPT? To save the NHS

Which patient has not consulted Dr Google? Indeed, which doctor, baffled by symptoms or keen just to confirm a hunch, has not turned on occasion to the internet? Today, however, a new technology is going further than hazy diagnosis, proving capable of analysing complex individual patient histories to suggest treatment plans as sophisticated as those of an expert medic. Ask nicely and it will even ditch the jargon and boil its recommendations down into a useful summary. For free. Turns out there is a new GP in town. 

Those letters stand not for General Practitioner, but Generative Pre-training, a method of getting computers to learn by combing through vast quantities of text, sound and images. Combine that with Transformer – a computational model that has radically improved the way machines make connections between sequential bits of information – and you have GPT, the digital world’s Big New Thing.

There is no business where the accurate distillation of information is as important as in healthcare, where all too often it is a matter of life and death. So could GPT come to the rescue of an ailing NHS? 

It certainly looks likely. Its latest iteration, GPT-4, has just been unveiled by San Francisco-based start-up OpenAI, demonstrating capabilities as limitless as those of the language it seems to understand and express with unnerving creativity and naturalness. 

For example, after signing up to try it on OpenAI’s site, I ask, “I’m feeling a little stiff with a headache and runny nose. What do you think might be the problem?” Less than a second later, GPT replies: “Based on the symptoms you described, it is possible that you have a common cold or a mild viral infection. Headache and runny nose are common symptoms of a cold, while feeling stiff may be a sign of general malaise or body aches associated with the infection.” It then advises that, “if symptoms persist” I should seek out a doctor. 

“Do you think your advice is as good as a doctor’s?” I ask. “As an AI language model, my advice is based solely on the information you provide,” it responds, “and it cannot replace the advice of a licensed medical professional.”

Can GPT relieve pressure on the NHS? I go on. “AI language models like GPT can be used to support the National Health Service in a variety of ways.” It lists several: automating routine tasks, analysing medical data, predicting patient outcomes, identifying potential areas of improvement in healthcare delivery, developing personalised treatment plans – all of which, it turns out, humans agree with. 

That sounds exciting, I think. But GPT is cautious. “The NHS has certainly faced challenges in implementing and harnessing the benefits of new technology in the past,” it notes. That, too, is true.

It sounds as though GPT truly understands me. Though of course, it doesn’t actually “understand” anything: what it is doing is predicting the most useful response to any question based on all that pre-training. But that does not stop it telling jokes to order, or building websites or composing novels in the style of your favourite author. Indeed, it excels at distilling huge amounts of information and responding with what can appear almost human insight. Certainly, when I ask GPT if we will cure cancer before 2050, its answer is detailed, perceptive and nuanced. 

Its potential is incredibly exciting but at the moment hard to quantify. “It has absolutely novel capabilities that are surprising even to experts,” says Alberto Favaro, director of healthcare at Faculty, the British AI firm that is currently working with NHS England to harness the technology. In particular, GPT’s creativity is helping the NHS to overcome privacy problems that prevent it mining its troves of data to design better treatments. 

Rather than combing through real, often fragmented, records, which must first be sifted to eliminate traces of personal data, the NHS is running a pilot using GPT software to generate “synthetic” histories that reflect reality but are not of real patients. These “ghost patients” allow it to run simulations modelling the impact of different treatment methods, working out which are most effective, without breaching privacy laws. “It’s like you’re asking it to write a novel about the patient,” says Favaro of the software. “It generates the data.” 

This creativity could also extend to clinical trials, he says. If GPT was used on the vast amount of existing clinical trial data, it could “suggest candidates for new drugs, or how new trials should be run”.

The process demonstrates one of GPT’s appeals to an NHS that has suffered repeated, and costly, technology failures: it is a tool to organise and draw conclusions from existing databases rather than an expensive new database itself. No wonder Dr Jonathan Pearson, NHS England’s lead data scientist, says the service is “excited to explore ways to use OpenAI technology safely and responsibly to support patient care”. 

Across the Atlantic, Anil Gehi has seen the power of GPT not just in data mining, but in clinical settings. The US cardiologist tested it by describing the history, symptoms and complications of a real hospital patient, deploying several technical terms, and asked GPT-4 for a treatment plan. It responded with what turned out to be the very plan Gehi had himself initiated. When Gehi tried other situations, he was again impressed. 

Perhaps that should come as no surprise. Google’s version of such a “Large Language Model” (LLM), as GPT-like software is known, scored 85 per cent when answering US medical licensing exam questions, well above the 60 per cent passing threshold for humans. Google acknowledged, however, that there is “a lot of work to be done” before it can be applied in real-world settings. This is because LLMs are still capable of getting the wrong end of the stick, or worse still, of what is known in the business as “hallucinating” – simply making stuff up. 

That hasn’t prevented doctors seeing their potential. Research shows LLMs are capable, for example, of sifting through medical records filled with jargon and abbreviations and extracting all the relevant information. In a recent study conducted by Massachusetts Institute of Technology, sentences such as “pt will dc vanco due to n/v” (the patient will discontinue the antibiotic vancomycin due to nausea/vomiting) were accurately deciphered 86 per cent of the time. 

For Dr Eric Topol, who led a NHS workforce review in 2018-19, the new technology will “drive keyboard liberation for clinicians”. “We’ll see… generative AI take on so many language-based tasks,” he says. “Office notes based on voice... aggregating and summarising a patient’s history from scouring their medical records.” 

Favaro suggests it is in such low-risk “paperwork” situations that GPT will break into healthcare. “Anyone knows that when you go to see your GP they are typing much of the time,” he says. “It’s obvious all that could be replaced by software, so that the GP can stop and listen to the patient.”

AI advocates believe it will free humans to focus on the stuff we’re good at: empathetic, interpersonal relationships. Understanding just from the way a patient walks into a surgery that there may be more to their symptoms. More, certainly, than a machine could pick up. 

At least for now. As Favaro says: “This is just the beginning.”