The Digital Doctor: Where Machines Excel at Diagnosis and Where They Mislead

A mole on the forearm: a doctor looks at it and reassures you; you show it to a digital system and it raises the alarm. Whom do you believe? The right question is not "which is cleverer?" but "for which task, on which patients, and with what safeguards?" Answering it shows where machines genuinely help us.

What the machine does well: one image, one task

The system trains on thousands of medical images labelled with their confirmed outcomes, extracts visual patterns, and turns each new image into a probability. It does not understand the disease; it is good at what it has seen often. The idea is set out in the article "What Is Artificial Intelligence, Really? A Plain Explanation Without the Hype".

There are results here worth pausing over. In a study published in Nature in 2020, researchers from the Google DeepMind team and their partners compared a mammogram-reading system with specialists' decisions. False alarms fell by 5.7 per cent and missed cases by 9.4 per cent (absolute figures) in the US sample, and by 1.2 and 2.7 per cent in the British sample. The system was trained on British data only and then tested on American data, where it kept its edge: reassuring evidence that it can generalise.

In 2018 the US Food and Drug Administration permitted marketing of IDx-DR, a system that examines retinal images in people with diabetes and gives a screening decision without a clinician interpreting the image. Its trial of 900 patients at ten primary-care sites showed a sensitivity of 87.4 per cent and a specificity of 89.5 per cent.

But the laboratory is not the clinic

Two reservations accompany those results. First, the mammography study was retrospective and some of its findings were simulations, and its authors acknowledged the need for prospective studies that follow real patients. Second, a model does not warn you when it steps outside its domain.

Skin is the clearest example. Researchers built a diverse set of biopsy-confirmed images and found that advanced AI models show marked shortcomings on dark skin and rare diseases, and that dermatologists themselves perform worse on those images. The good news is that training the models on diverse images closed the gap between skin tones. Fairness in diagnosis begins with fairness of representation in the data, as the article "Algorithmic Bias" explains.

Doctor plus machine: pairing is not guaranteed

Many assume that a doctor assisted by a machine inevitably beats either alone. The evidence is subtler. In a randomised trial of 50 physicians (2024), access to a large language model improved diagnostic reasoning by only two percentage points, not statistically significant (76 per cent against 74 per cent), while the model alone outperformed both groups. The manner of combining, not the mere presence of the tool, is what matters.

There is a further, quieter risk: erosion of skill. In an observational study at four Polish centres (2025), doctors' adenoma detection rate in colonoscopies performed without assistance fell from 28.4 to 22.4 per cent after they had become used to an automated detection system. It is observational and cannot prove causation, and the period was short, but it warns that reliance may blunt a doctor's instinct if it is not safeguarded.

What to do

• The doctor decides: the system remains an aid, and responsibility lies with the human who signs.
• Local validation: no system is adopted before being tested on the region's own patients, with their data and equipment, rather than on figures announced elsewhere.
• Preserve skill: train doctors in the tool's limits and in reviewing its decisions rather than accepting what it displays.
• Inform the patient: a patient may know that an automated system took part in the diagnosis, and may ask and object.

In the Kingdom, the Saudi Food and Drug Authority issued in 2023 a guidance document specific to AI- and machine-learning-enabled medical devices, a useful regulatory framework. The more diverse Arabic health data we build and protect, the more the "digital doctor" becomes our doctor. The aim in the end is a person who is reassured and healed, God willing, not a number that climbs in accuracy on a table.