What AI’s potential to improve accurate diagnosis means for the health system

Five years ago, AI was struggling to identify cats. Now, it can interpret medical images as well as a team of experts with decades of combined training. More accurate diagnosis can reduce misdiagnosis, and reduce the need for second opinions and follow-up tests. This has tremendous benefits for the system as a whole.

By Adrian Baker

Read time: 8 mins

Image credit: Harlie Raethel on Unsplash

Share article

If the health system is a bit like a road network, then medical diagnostics are a bit like spot checks along this road network – slowing things down whilst cars are examined individually for problems. But imagine, instead of individual people checking each vehicle, cars just move freely thanks to a smart camera system. No more frustration, no more traffic – and you’ll get to where you need to be a lot quicker. This is the potential of what AI-powered diagnostics can bring to the health system.

The situation (but only for now)

Diagnosis can be a bottleneck in the flow of patients in the health system.  One of the reasons for this is that the demand for diagnosis is higher than the available resources.  Not enough MRI machines, not enough CT scanners, not enough staff to look at the scans, and not enough hours.  And as the prevalence of conditions that require medical imaging and blood samples increases, such as cancer or arthritis, diagnosis is put under increasing strain. 

So when we read stories of another AI algorithm beating radiologists at diagnosing medical images of yet another condition, it’s natural to feel a sense of excitement.  But critics will claim that AI imaging diagnosis is still fallible.  In one Stanford University study, researchers trained an AI algorithm to detect abnormalities in over 40,000 radiograph images.  The findings were mixed – the AI performed better than radiologists at detecting abnormalities in images of fingers, and about the same as radiologists at detecting abnormalities in wrists.  But, in elbows, forearms, hands, and shoulders, the radiologists performed better.   

But there is no denying just how far AI imaging has come from the days of struggling with cat pictures.  And it should be noted that trials of AI vs radiologists tend to compare the AI to the majority performance of a group of radiologists.  In that sense, not all radiologists get it right all the time.  In another study out of Stanford, researchers developed an algorithm called ChexNet to diagnose pneumonia in chest X-rays.  Pneumonia can be challenging to diagnose, especially from chest X-rays alone, and there can be a lot of variation between radiologists in terms of diagnosis.  But in this study, ChexNet performed better than the radiologists.  So whilst the current performance of AI compared to radiologists is mixed, the trend will only go in one direction.

Going beyond individual cases

The reason AI driven diagnosis is such a breakthrough is less to do with the technology and more to do with the impact on the system of dramatically improving diagnostic accuracy. 

In a collaboration between MIT and Massachusetts General Hospital, researchers trained a machine learning algorithm on high-risk breast lesions.  Not all high-risk lesions develop into cancer: some lesions carry a 20% risk and require surgical removal whilst others just require close surveying.  The AI trained by the researchers was able to identify 97% of lesions that turned into cancer, meaning that it would have helped avoid 30% of unnecessary surgeries.  Surgery itself carries risk and unwanted costs, so avoiding unnecessary surgical removal is always beneficial. 

For an individual, avoiding unnecessary surgery has obvious psychological and physical health benefits.  But on a wider scale, a 30% reduction in surgeries would mean fewer hospital beds being taken up, freeing time for doctors and nurses.  In turn, more time for doctors and nurses may mean more attention is given to other cases, improving clinical outcomes and care in other areas.  More time may also mean a reduction in stress and pressure for the staff, reducing errors and sickness absence and saving the hospital money that they could also spend in other areas.  For the 30% of people who avoided the surgery, it means less time away from work, for themselves and perhaps for their family who might have to care for them in recovery.  Of course, this is an idealised scenario that depends on these potential benefits being taken advantage of at every step, but it also represents an opportunity for health systems to finally take control of the challenges created by changing patient demographics. 

Reducing second opinions, follow up tests and misdiagnoses

Being more accurate means eliminating the need for follow-up tests and second opinions.  This in turn has a multiplier effect because it reduces the time to diagnose a particular condition.  Reducing the waiting time for diagnosis improves the chance of survival for conditions such as cancer, and that has benefits which are so obvious it seems almost crude to describe in the form of a benefits analysis. 

Reducing the need for second opinions and follow up tests also means possible reductions in the amount of treatment needed.  From the relatively simple to the more serious, it is an infallible truth that getting to conditions early means avoiding complications and more intensive treatments.  Reducing the amount of complicated or expensive treatment needed also frees up resources in hospitals, again, saving costs and allowing that extra time or investment to go to other areas.  Similarly, it reduces the costs incurred by patients, and these can extend far beyond the direct costs of the drugs themselves; with travel costs to hospitals, lost income from work, and a reduction in the quality of life all stemming from the impact of delayed treatment.   

Improved accuracy also means reducing misdiagnosis.  Whilst usually an initial misdiagnosis gets rectified, it only does so after delays and other associated costs.  But in the most severe cases misdiagnosis can lead to irreversible complications or death, so eliminating or significantly reducing misdiagnosis would take away suffering of errors that should never occur.  And if that isn’t enough, reducing misdiagnosis would help put a dent in the staggering costs of litigation.  AI therefore represents an opportunity to stop, or drastically reduce diagnostic errors; errors which should never occur, errors which stop time and money being spent on taking healthcare to the next level, and errors that for the individuals affected by them can leave a mental scar long after the tissue heals.

What’s possible vs. what’s implementable

There’s no denying the speed at which AI is being trained.  For example, the researchers behind ChexNet took merely a month to train their algorithm.  That’s one month compared to the years upon years of training.  And static images of X-Rays, CT scans and MRIs aren’t the only imaging diagnosis that can be aided by AI.  For example, a team at Showa University, Japan used AI-assisted endoscopy to assess colorectal polyps in real-time.  The real-time AI diagnosis of the video feed helped reduce unnecessary polypectomy, reducing the burden on diagnosing biopsies.  But there is a chasm, carved by reality and historical structures, that means ‘what is possible’ is almost improbable to reach. 

There are those that believe AI will replace radiologists.  But all it would take is one or two misdiagnoses and that will grind to a halt any move towards a completely autonomous diagnosis system.  It wouldn’t matter that healthcare professionals make errors (medical errors are the 3rd leading cause of death in the US), so long as it’s healthcare professionals and not AI that make them. 

So the likelihood is that we will move towards a symbiotic relationship with AI, and that is no bad thing.  AI diagnosis will likely be a tool that’s used to augment the work of radiologists, increasing speed, accuracy, and productivity tantamount or even greater to the introduction of computers.  The likely scenario will be that pressures placed on the health system will force hospitals to use autonomous AI to diagnose medical images in cases where there’s a very high degree of confidence.  And that is a much easier case to make to the public: radiologists shouldn’t be spending their time looking at thousands upon thousands of non-urgent, obvious, medical images.   

Unblocking the whole system

To actually ensure that the benefits of AI-driven diagnosis are experienced throughout the whole health system, a number of obstacles need to be overcome.  Achieving the productivity utopia mentioned takes more than installing a modern IT infrastructure and equipping a hospital with AI-driven diagnostics.  That’s challenging enough, and doesn’t even take into account human behavioural factors such as resistance to change, lack of training in using new technologies or how data should be handled. Taking the road-network analogy a step further, there is no point in relieving a traffic bottleneck in one part of the system, only for the traffic to grind to a halt a few meters down the road. 

The improved accuracy that AI diagnosis can bring will certainly help reduce costs and inefficiencies over the long-term, but to maximise the potential benefits, whole patient pathways would need to be looked at.  The adage that you are only as slow as the slowest member in your group would therefore ring true.  Diagnostics might be sped up, but patients might still have to wait for machines to be free (although reducing the need to re-tests and second opinions would help with this), there may still be shortages of beds, or delays in primary and social care.  All these pressure points along the pathway require unclogging.  No matter how effective AI might be in helping doctors diagnose patients, the system-wide benefits will inevitably fall on human factors.

Adrian Baker

About the author

I look at emerging technologies from a social science and policy perspective. I completed my PhD on the diffusion of innovations at University College London, looking at how innovation gets implemented in healthcare organisations. My main interests are on policies that encourage innovation and the diffusion of emerging technologies, and understanding their social implications so that everybody wins.