杏吧原创

HOW would you feel if your doctor, rather than consult their own clinical knowledge, turned instead to an AI trained on your medical history to help diagnose your next ailment or write your next prescription?

These sorts of scenarios have been hypothetical for decades 鈥� the technology has been subpar and the stakes too high to risk offloading medical advice to a machine. However, the success of large language models like ChatGPT, a popular, artificially intelligent chatbot from the OpenAI research lab, has led to a rethink of what might be possible.

In December, I was reading through a list of machine learning preprints 鈥� scientific papers that have yet to undergo peer review 鈥� when I came across , a medical machine learning model from researchers at King鈥檚 College London (KCL). It uses GPT-3, the model that powers ChatGPT, and real electronic health records (EHRs) to predict 鈥渇uture events, estimate risk, suggest alternative diagnoses or forecast complications鈥� for simulated or real people whose information is fed into it. While intriguing, I would have paid less attention if it wasn鈥檛 for a personal connection.

King鈥檚 is my alma mater and the model had been trained on real-world, anonymised data from King鈥檚 College Hospital from 2010 to 2019. I had visited its emergency departments at least twice during that period while studying. I soon found Foresight wasn鈥檛 the only medical AI. At the end of 2022, Google announced , a version of its enormous general purpose AI model PaLM, which is trained on text from the web and books and fine-tuned using medical documents.

Some of Google鈥檚 claims were remarkable: its AI could answer common medical questions requiring a long written response, and real doctors said 92.6 per cent of Med-PaLM鈥檚 replies 鈥渁ligned with scientific consensus鈥�, just 0.3 per cent less than answers given by human doctors. While the medics assessing its abilities noted that gaps in some answers and possible safety issues meant the model wasn鈥檛 yet suitable for clinical use, it was clear such AIs were making rapid progress.

It looks like medical AI models might achieve a clinical level of competency before regulatory bodies catch up. All this left me wondering how close we are to seeing these tools in clinics and whether my own medical data helped train an AI (and if it was truly anonymised). I spoke with two of Foresight鈥檚 creators to get some clarity. Yes, if I had visited a King鈥檚 College Hospital A&E during that period, then my data would have been used to train the model, at KCL told me. And it was almost certainly anonymised, he said.

The EHRs had had any potentially identifying information removed, like rare diseases where there were less than 100 samples, and you couldn鈥檛 ultimately get patient-level data out of the AI system. 鈥淭he risk of re-identification within the model is effectively zero,鈥� says Foresight team member James Teo. I was reassured, but still a little spooked at the thought of a digital twin within the model.

Those behind Foresight, as with Med-PaLM, have (for now) opted not to use the AI in a clinical setting and so it won鈥檛 be assisting medics yet. But it, too, is producing encouraging : five doctors assessed its predictions for future health issues for 34 simulated patients and found its top forecasted condition was relevant 97 per cent of the time.

Teo couldn鈥檛 give me an exact date on when Foresight might be ready for real-world use and said that they needed another year or so to collect data on the model鈥檚 accuracy and explainability. But it looks like the models might achieve a clinical level of competency before bodies like the UK鈥檚 Medicines and Healthcare Products Regulatory Agency (MHRA) are in a position to properly assess them, he added. 鈥淭o a certain degree, [Foresight] reaches technical feasibility,鈥� says Teo. 鈥淲hether it meets regulatory feasibility still requires the regulators to develop maturity in their frameworks.鈥�

We don鈥檛 know what those frameworks might be, but they are likely to look at whether the AIs can reliably produce accurate answers and the transparency of their decision-making. AIs will also need to show they aren鈥檛 biased towards certain groups of people, a risk for machine learning models because of the way they are trained. That is especially pertinent in healthcare, where demographics can affect medical outcomes, says at The Alan Turing Institute, UK.

Another pitfall could be so-called AI hallucinations, where the system erroneously generates 鈥渇antastical, unfaithful, or nonsensical outputs鈥�, says Leslie.

All these will be questions for the MHRA when the time comes, but, if the preliminary results from these models are improved upon, it will be a matter of when, not if, your next diagnosis is AI assisted.

Alex Wilkins is a New 杏吧原创听reporter who covers听artificial intelligence, physics听and space. 鈥淎rtificially intelligent鈥� is a column that cuts through the hype, looks at what AI is really capable of and what it means听for us. You can follow him @AlexWilkins22