19 February 2025

AI in breast cancer screening may do more harm than good

The complexity of medical practice is at odds with politicians’ thirst for good headlines.

The EDITH trial – announced by the government to coincide with World Cancer Day on 4 February – aims to recruit 700,000 women to investigate the use of AI in the breast cancer screening programme. Currently, every mammogram (X-rays designed to pick up signs of cancer) is read by two radiology specialists to minimise the chance of mistakes. The Early Detection using Information Technology in Health trial will replace one of these human doctors with an AI diagnostic tool.

Participants will be recruited from women “already booked in for routine screenings” – one hopes they will be sent literature in advance, so they can consider it properly. And trial information needs to be presented in a balanced and neutral way. The Department of Health and Social Care’s (DHSC) press release does not augur well in this respect: it is peppered with superlatives, citing the “cutting-edge” technology which will help “radiologists catch breast cancer earlier”, part of the government’s “unleashing” of AI to “drive up health services”. According to the DHSC, the AI tool enables just one radiologist “to complete the same mammogram screening process safely and efficiently”. If we know that already then there is no point conducting the trial.

Ironically, the potential for harm does not lie in missed cases. Most research over the past few years suggests that AI – though not infallible itself – can outperform even two radiologists at picking up suspicious changes. Nor does it appear to increase the rate of “false positives” – the substantial numbers who are recalled for further tests following mammography and who will, after an often agonising wait, ultimately be given the all-clear. Where AI might worsen outcomes lies in an under-discussed aspect of the current screening programme, which is that it doesn’t “save” any lives at all and actually causes net harm.

Mammography does make a small difference to breast cancer mortality: for 1,000 women screened for a decade, one will avoid dying from the disease. This is the root of the DHSC’s estimate that the UK’s screening programme “saves” 1,300 lives a year. Yet for every woman who benefits, another three are diagnosed with a breast cancer that would never have caused them any harm – a phenomenon known as overdiagnosis. The last time the UK’s screening programme was evaluated in 2012, this awkward fact – that screening turns a substantial number of healthy women unnecessarily into cancer patients – was deemed an acceptable price for the reduction in mortality from the disease.

Yet this conclusion failed to take account of twin phenomena. First, treatment of breast cancer has advanced hugely over the past 40 years. Most women presenting with symptomatic disease today can expect to survive, meaning that the benefits suggested by the early trials of screening have been greatly attenuated. Second, we still have no way of determining which screen-detected tumours are overdiagnosed, so all are treated. Surgery and radiotherapy both cause substantial harm and occasionally death. The net result of these phenomena is to cancel out any survival advantage from screening, something that has led Professor Michael Baum, one of the architects of the UK screening programme, to call for its cessation.

The complexities and nuances of medical practice are fundamentally at odds with politicians’ thirst for simplistic good-news headlines. The quest to incorporate AI into the breast screening programme is being driven by the productivity gains to be had from replacing human doctors. But an arguably more important use of the £11m funding would have been to research use of AI to reduce overdiagnosis. AI’s ability to discern patterns on mammograms or biopsy specimens that might identify indolent tumours not requiring treatment, or in risk-stratifying personal characteristics that might affect prognosis, could potentially restore value to a breast screening programme that can be injurious to participate in.

[See also: James Blunt live: a nostalgia karaoke]