Constant cough? Abdominal pains? Change in bowel habits? These are the types of problems any healthcare system should be able to help you with.
It should be simple: the doctor sees you, arranges scans and blood tests, you’re given the results, then the diagnosis and treatment — and hopefully you’re cured.
Of course, it isn’t that easy, not least because there just aren’t enough doctors to see patients in a timely manner.
But also, crucially, we as doctors aren’t actually as good as we think when it comes to making the correct diagnosis. The average misdiagnosis rate is 9.7 per cent, according to a major review of the evidence, published in the journal Diagnosis in 2020.
That’s where artificial intelligence (AI) can step in.
The average misdiagnosis rate is 9.7 per cent, according to a major review of the evidence, published in the journal Diagnosis in 2020. That’s where artificial intelligence (AI) can step in (File image)
There has been a lot about it in the Press recently, with dire warnings that, if uncontrolled, AI threatens the survival of the human race.
But when it comes to medicine and, specifically, using it for diagnosis, I’m all for it. There’s a growing body of evidence showing that AI can help.
A study last year in the Korean Journal of Radiology found that AI can significantly improve the ability of radiologists to pick up breast cancers from mammograms in screening programmes. This would not only improve outcomes for patients but also save money and time — and help make jobs such as mine more manageable.
I know there are plenty of patients who have had to wait too long to see me, and too long to get scan results back. And, like all medics, I have also been wrong about some diagnoses.
However, I would never want to be treated by an AI doctor, as some are suggesting: the idea of seeing only a robot with artificial intelligence at your time of need is both abhorrent and dangerous.
As a medic, you often rely on human instinct to find out from the patient what’s really going on — you can’t just base your interpretation on what they initially say to you.
A study last year in the Korean Journal of Radiology found that AI can significantly improve the ability of radiologists to pick up breast cancers from mammograms in screening programmes (File image)
A number of years ago, I remember seeing a teenage girl with knee pain. Initially, I couldn’t work out what had caused her pain because she denied every possible suggestion about what might have triggered it, from a sports injury to getting drunk and not remembering an accident.
She seemed incredibly distressed by the pain and I had a gut feeling that all was not right.
After about 20 minutes, sensing I had earned her trust, I asked her parents to leave the room. That’s when she confided that she’d had unprotected sexual intercourse, and now had an unpleasant discharge.
I put two and two together and worked out that she had gonococcal arthritis — an infection of her knee joint caused by gonorrhoea.
The treatment was a course of antibiotics but, crucially, it also involved careful handling of the situation with her parents.
A computer would never have earned that young girl’s trust to reveal the information that led to the diagnosis — let alone been able to handle this matter delicately with her family.
AI can also never replace humans when it comes to holding a patient’s hand, listening to their wishes and deciding that, rather than treating a cancer, actually the best course of action may be to make them comfortable in their last days. There is much more to being a good doctor than simply giving the right diagnosis.
But where AI does have the potential to help is in making that diagnosis accurate.
Essentially, there are two types of patients. First, there are those whose symptoms fit a pattern we recognise, so we quickly jump to a conclusion (often before or without test results); then there are the patients whose cases we have to ponder.
Diagnosing the first is fraught with unconscious bias — such as when a condition is at the front of your mind and you don’t consider other possible explanations. I fell into that mistake in the height of the second wave of Covid when I initially missed a very nasty infection in a 25-year-old who had a rare complication of bacterial tonsillitis. I had quickly assumed his symptoms were due to coronavirus.
Or sometimes you’re swayed by what another healthcare worker says. For example, I’ve seen a patient with slurred speech who was misdiagnosed by the initial doctor as being intoxicated because that’s what the paramedics had thought.
With time — and a CT scan — I diagnosed that in fact the patient had had a stroke.
This is where diagnostic AI can help. Take eyes: research has shown that two-thirds of patients with acute eye conditions are misdiagnosed when ‘triaged’ by front-line practitioners who aren’t eye specialists.
So there’s now a trial, funded by the National Institute for Health and Care Research at the A&E of Moorfields Eye Hospital in London, which is looking at using AI to improve this and free up specialists’ time. The AI (built with world-leading doctors) takes all the information from the nurse or optometrist who has assessed the patient, and then tells them whether a patient needs urgent attention from a doctor or a less urgent referral.
While the results are not yet published, the lead researcher, associate professor Alex Day, told me that this ‘has the potential to help our patients to be treated as quickly as possible, giving them a better outcome and experience’.
But I’m most excited about AI when it comes to those patients where we have time to think and ponder — and the biggest game-changer is how AI can improve the accuracy of tests. Or rather, the accuracy of the interpretation of the results.
For instance, CT scans are incredibly hard to interpret, which is why you need 15 years’ training to become a consultant radiologist. But humans are not always perfect. Could artificial intelligence help? The answer, increasingly, is yes.
One of the best examples is the new NHS lung cancer screening programme for smokers. The AI used in the scans automatically measures the size of any nodules (small lumps) in the lungs, which can be the first sign of the disease, flagging them up to the radiologist to interpret.
This improved the accuracy of the radiologist’s diagnosis, reported the European Journal of Radiology last year. It also cut the time the radiologist took to assess and write their report on the scan by between 33 and 44 per cent. This was particularly significant as there’s a national shortage of radiologists.
The next logical step is to see whether AI completely negates the need for a human radiologist. When radiologists were pitted against stand-alone AI systems in assessing scans of people with lung cancer, the average radiologist’s interpretation was accurate 91.7 per cent of the time, while the best-performing stand-alone AI was 90.9 per cent, reported the journal Radiology: Artificial Intelligence in 2021.
So, at the moment, AI should be used to help radiologists interpret scans, but in time could replace them in this role.
Another way AI could help improve doctors’ analysis of test results is when there are multiple factors at play.
For example, if you suspect a patient has had a heart attack, to diagnose it you have to take into consideration their story and their risk factors, the results of heart scans and blood tests — most importantly, the presence of a protein, troponin, in their blood.
Troponin can be a sign of heart damage, but you can also see it for other reasons, such as after exercise or with kidney failure.
So we use a cut-off level — and patients whose levels are above this threshold have had a heart attack, and those below haven’t. But this means we will miss some heart attacks.
But now, AI software can help work out the probability of it being a heart attack based on the risk factors and an examination and ECG reading before the blood test, and adjust the cut-off based on that probability, reported the journal Nature Medicine a few months ago.
As a result, doctors would be able to more accurately work out which patients are fine and can be discharged, and those who need to be admitted.
The bottom line is this: cautiously and only where there is the evidence to do so, we need to adopt computer AI to help improve the way that doctors make diagnoses.
But, rest assured, I can never see the day when they will replace actual human doctors.
@drrobgalloway
For all the latest health News Click Here
