Data diagnostics: the role of AI in clinical diagnosis

Diagnosis of some of Scotland's most deadly illnesses is not an easy task. Acute heart failure, Scotland’s biggest killer, for example, is very difficult to accurately diagnose given patients’ present with a variety of symptoms. While a disease like bowel cancer, the second most deadly cancer in Scotland, can be a lengthy, uncomfortable and intrusive process to have confirmed. There is no magic bullet to prevent or cure these illnesses, but new technology is helping to diagnose far quicker and personalise patient care.

One of the potentially crucial new technologies that could promote personalised patient care, improve the accuracy of diagnosis, and save time, while bringing down costs is machine learning – artificial intelligence (AI).

Although some may find AI difficult to understand and the concept futuristic, it is all around us in our everyday life, and it has been for years. Unlocking your smartphone with facial recognition, simply scrolling social media, using websites that make recommendations like Amazon and Netflix, or using a GPS like Google Maps, the average Scot is frequently interacting with some form of it every day.

And NHS Scotland has already made inroads to incorporate effective AI software into its service. Last year NHS Golden Jubilee in Clydebank became the first hospital in Scotland to use AI to assist in the installation of stents in arteries near the heart. The Ultreon 1.0 software, designed by Abbott, is linked to an Optical Coherence Tomography (OCT) device, which uses infrared light to look at the calcification in a patient’s artery.

The AI component in the software, which is used on 20 per cent of patients across the hospital’s four coronary catheter labs, can measure very accurately and very quickly the thickness of the calcium, and how circumferential it is. The algorithm is then used to assist the clinician in their decision-making; whether to use a shockwave balloon or a Rotational Atherectomy kit to modify that calcium.

Dr Stuart Watkins, consultant interventional cardiologist at NHS Golden Jubilee explains how the new AI software improves the preparation before installing a stent: “It helps refine the procedure, telling you how bad the calcification is, so we then know what to do to the artery before we stent it.

“It also is a very quick way of measuring how long the narrowing is and how long the stents are. “Before, we would pass the catheter down the artery and the run would be a couple of minutes, then you would stop, go offline and do line measurements on the software to measure how thick it was and look at it, whereas now that is all done instantaneously pretty much. You can flick through and see the highlighted areas of concern.

“It cuts out maybe a few minutes, but if you do this a lot in patients and you are doing three or four runs in the same patient it does reduce the time.

“The fact that all the numbers are there, you are not having to measure them, is very helpful to us. It speeds up the process. It is good for patients; it helps us look at things that we wouldn’t normally look at. Because it has picked up on something that we haven’t picked up on. I think it is very clever and it is going to be the way forward.”

Although some inroads have been made, AI software isn’t something that has been implemented widely in NHS Scotland. But that could be about to change. Glasgow and Edinburgh University researchers are currently developing AI systems that will improve the diagnostic accuracy of life-threatening illnesses like acute heart failure and bowel cancer.

Syed Ahmar Shah, a chancellor’s fellow at the Edinburgh Medical School describes the current healthcare system as “reactive”. He believes adoption of AI in medical practice should be more common: “Data is crucial, it’s going to be at the heart of this [transformation], but it alone is not sufficient. You need the technology around it; enabling technologies where you can exploit this data.

“The current healthcare system, the model itself, it’s not sustainable. So, we need a more proactive approach where we can focus more on the early stages so that fewer and fewer people go end up in hospitals.

“We saw this in Covid. It was the first pandemic in the age of data and AI. There were things that were done quite rapidly, and the UK was leading in a lot of this.”

Nearly one million people in the UK are living with heart failure. That is expected to rise by 50 per cent in the next 25 years. While it is rather common, it remains difficult to diagnose. Dr Ken Lee, a cardiology specialist registrar and clinical lecturer at the University of Edinburgh helped develop CoDE-HF, a diagnostic tool that can be used to determine the probability of acute heart failure with the help of AI.

Lee explains why it is so important to get an early diagnosis: “We have very good evidence that if we diagnose these patients early, we start treatment early, patients do a lot better. But the problem is most patients with heart failure come into the hospital with breathlessness, which is a very common symptom. We must consider many other potential conditions, like pneumonia and anaemia, for example, which are also quite common medical conditions.”

The difficulty to get an accurate diagnosis led to the development of CoDE-HF, which was theorised to be far more accurate at picking out acute heart failure in patients, particularly those with existing comorbidities. After analysing the largest ever dataset on the current blood test used to diagnose acute heart failure, Lee and his colleagues found that there were flaws in the blood test. Patients with previous pre-existing heart failure and renal problems could be missed at anywhere from “one in five patients to one in 20 patients.”

Lee and his colleagues understood the opportunity to apply AI in the diagnosis stage of acute heart failure. The team had a good starting point as they already knew “about the challenges in diagnosing heart failure, the biomarkers and clinical practice". Using that information, they incorporated all the factors known about heart failure into AI algorithms to produce an individualised probability of acute heart failure for each patient. The proof of concept allows clinicians to make far more personalised decisions for their patients, allowing them to better serve patients with complex medical problems that often coexist with acute heart failure.  

Lee explains the new diagnostic tool: “This is about being able to diagnose heart failure, acutely and promptly. Patients with heart failure are increasingly more elderly and have multiple other medical problems like  diabetes and kidney problems, making it rather complex to make a diagnosis. When they come in, they’re breathless, and sometimes they can be very sick.

“As a doctor, I ask myself what could this be? There is a list of what we call differential diagnosis. Even in my head, I am thinking could it be pneumonia, COPD, or anaemia? It can be quite challenging to diagnose this type of patient. So, there may be quite a delay before they receive the treatment they need. 

“We have developed a prototype app, which is deployed in a smartphone. You key in the age, renal function, the different conditions, and the blood tests and press enter, and the algorithm runs. Within a split second, it tells you the percentage chance that it could be acute heart failure. Then you can refer the patient to cardiologists to get a specialist opinion if heart failure is likely.”

A pre-pandemic study, involving 13 countries and 10,369 patients looked at the accuracy of diagnosis compared with the standard NTproBNP blood test. In every metric, CoDE-HF performed better than the standard test done in NHS Scotland. Now researchers want to know if this test still works post-covid.

Glasgow University’s INCISE project is another aiming to improve diagnosis, by using AI in bowel cancer screenings. They have developed a tool that can predict which patients with precancerous growths in their bowels, called polyps, will develop further polyps later in life.

Patients who test positive for blood in their stool are invited for a colonoscopy. Five per cent of patients are found to have cancer, while a further 30 per cent will have polyps. The polyps are removed, however around half will develop new polyps. By creating a risk stratification tool, which predicts polyp reoccurrence, INCISE hopes that they will cut in half the number of people that need to go back for the uncomfortable reassessment, allowing NHS Scotland to reduce their surveillance list and allowing resources to be focused on higher risk patients.

Professor Joanne Edwards, director of INCISE, explains how it will work: “We have a patient cohort of 2,600 that had their first polyp when they first came for a colonoscopy, we know what has happened to them over five years.

“We can use their first polyp to predict whether they will develop a new polyp or not. We are taking all their clinical information and doing molecular analysis of the polyps to look at how the piece of tissue changes, looking for certain markers that we know are associated with high risk colorectal cancer, along with genomic and transcriptomic analysis.

“We are putting all of this data together and doing machine learning on the data to develop a novel algorithm that can tell us what pieces of information are important and helps us betterpredict the patients that will need to come back for a surveillance scope.”

The AI component to this research is crucial, without it “[diagnosis] would take much longer, as you would have to analyse all of the individual components” and it will also shorten the time to diagnosis as doctors will interact with “a portal that the information will be fed into so that the doctor does not need to understand in great detail how the risk has been calculated”.

Now that they have the data INCISE is working to define the risk score before they try the new diagnostic tool on an independent cohort of patients.

INCISE project manager Gerrard Lynch pointed out how this software could solve an issue that is particular to the west of Scotland. He says: “We do have a problem, particularly in the west of Scotland, of patients missing their surveillance scopes because perhaps they don’t think they are truly high risk, and if you are self-employed, it has an economic impact on you.

“To have a much more detailed risk score, it will make the decision-making a whole lot more informed.”

Lee believes that in Scotland we are best placed to carry out more AI-assisted software for medical care, as we transform our healthcare system: “This is probably one of the best places in the world to do data-driven innovation because of our universal healthcare system. Whereas if you do the same research in countries where healthcare is primarily provided by the private sector, you get quite a skewed view of the situation because to qualify for a certain health plan you need to be of certain socioeconomic status.

“In the NHS, everyone has equal access to healthcare regardless of ethnicity, socio-economic status, background. So, you have this huge resource of healthcare data to be able to do this type of research and we can be confident that the findings are generalisable to the broader population.

“The availability of data is unique to Scotland because in many places in the world there just isn’t electronic healthcare data infrastructure to be able to do this type of research, and to build good quality algorithms you need good quality data to train and develop the algorithm.”