The future will see you now

13 minute read

AI will revolutionise rheumatology imaging and diagnostics. But what are the risks?

Artificial intelligence (AI) is making inroads across most fields of medicine, and its inevitable use in rheumatology heralds the potential for more widespread access and reduced workloads for specialists, along with improved screening, diagnosis and treatment for patients.

However, the use of AI in clinical practice also raises many legitimate concerns, from the risks of relying on an often-opaque decision-making process, to the potential for AI to magnify incorrect assumptions and biases, and the concerns of patients about its safety.

Rheumatologist and ARA spokesperson on technology, Professor Rebecca Grainger, says that while many clinicians may be understandably apprehensive about the introduction of AI, she sees it as another very useful tool to help understand disease.

She says that AI’s role in practice will provide similar utility to that of peer-reviewed clinical studies.

“Previously, we’ve used statistical methods in academic studies to help us understand patterns and associations around disease, and to consider how the findings might apply to our own setting.

“To interpret conventional studies, we ask questions like: who is the population that the study is sampling from? Are there any problems with the samples? What are the limits of the statistical tools that are used, and how reliable and applicable are the findings?”

This has meant doctors need an understanding of both the disease, and the mathematical process used to interpret findings from various studies, Professor Grainger says.

“The same applies to AI and machine learning. You can’t just randomly choose a sample, create a model and then apply it to other settings unless you understand the data that the model has been trained on, and whether the application of this model is appropriate to the setting.”

AI and deep learning

Melbourne rheumatologist Dr Christopher McMaster is immersed in a range of rheumatology-related research projects that use data science methodologies.

These include using algorithms to analyse PET scans with colleagues at Austin Health, and applying machine learning to do radiographic scoring of x-rays with an Adelaide-based colleague.

Dr McMaster worked as a data scientist before studying medicine, and recently published an overview of the application of AI in rheumatology [1].

He says that rheumatology is rife with “unstructured” data, such as images and free text, which can be difficult to analyse at scale – but AI will let us to unlock a wealth of information held within these data formats.

The emergence of deep learning, where AI can “learn” the structure underlying collections of images and text, allows AI programs to improve their responses over time as they gather more data.

Rheumatology has small datasets compared to many other specialties, Dr McMaster says, but the diagnostic labels within various conditions can be far more difficult to categorise.

“For example, even though we have classification criteria for diagnoses, assessing things like the amount of bone erosion or the extent of connective tissue disease can involve a level of subjectivity and possible inter-observer disagreement,” he says.

Pathophysiology underlies all disease, he adds, but no two people with the same classification or diagnostic label will have similar immunologic mechanisms, so a lot of heterogeneity exists even within the one label.

“While that can initially make it hard to apply AI, it also makes it really promising because deep learning can absorb a whole lot of information from images and data, and then identify connections and subtle features to either differentiate or associate these, in ways that humans find difficult,” he says.

AI can process more information

Dr McMaster says that because clinicians can only see a finite number of patients, their practical knowledge and exposure to various conditions is limited to the patients who present to them, which can be somewhat random.

“With rarer disease, you may only see a certain condition very occasionally, so you’re only seeing a small sample, separated in time, making it hard to connect all the dots,” he says.

In comparison, a deep learning model which gathers data from multiple centres will, by sheer weight of numbers, have a less biased sample.

“Supervised learning techniques are very effective for diseases that have discrete labels, for example, to distinguish rheumatoid arthritis from osteoarthritis,” he says.

However, these techniques fall down in the face of disease where there’s very diverse presentations, such as lupus, he says, where different patients might respond to certain treatments differently.

“This is where deep learning, or unsupervised learning, is useful, as it takes a data-driven approach to classification,” Dr McMaster says.

“This could be an advantage over current methods where we clinicians, from our limited sample and our own biases, decide the criteria and methods for classifying patients,” he says.

“Data-driven identification of common features can tap into a far larger information base to characterise disease.”

Genius methods to train AI models

While some specialties grapple with huge volumes of data, getting the most out of rheumatology’s smaller data sets has involved some clever techniques.

AI models can be trained using “contrast learning”: giving it random images and training the model to tell them apart.

“This helps an AI model learn general features which can then be used and transported to your dataset,” Dr McMaster says.

Other methods include augmentation – where random alterations are introduced to the small data set.

“The model is given a few different copies of each image to learn from, with random alterations each time that reflect possible variations – the image might be rotated, flipped, or have some noise introduced,” Dr McMaster explains.

More useful are the different ways to gather real-world data from a range of sources.

“Taking anonymised images from a group of hospitals and pooling their data sets allows us to improve pattern recognition, and certain private radiology services may also allow access to much larger data sets,” he says.

Another tool called federated learning involves multiple sites that contribute to training a model, not by sharing their data, but by sharing the changes to this model, which is distributed among different sites to each train on their own dataset.

Rich sources of data

Unstructured data in rheumatology can come from a range of sites including the text in biopsy reports, electronic health records and images such as ultrasounds and radiographs.

Rapid advances in natural language processing (NLP), or the ability for computers to identify important keywords and patterns from unstructured text, means that AI models can extract surprising amounts of information from clinical notes.

One example is the use of NLP to identify patients with rheumatic diseases from electronic health records, which Dr McMaster and colleagues trialled on a series of temporal artery biopsy reports, with results presented at ACR Convergence in 2021[2].

Their AI model classified temporal artery biopsy reports based on the presence of three histopathological features (adventitial inflammation, giant cells, intimal hyperplasia) and the overall conclusion (giant cell arteritis or not).

It was trained on 161 biopsy reports from one centre, then tested on 220 biopsy reports from a second centre.  Initial results showed very high accuracy of 99% from the original centre, and 93% from the second centre.

Dr McMaster says a decision about the desired outcome must be made up-front before training AI models.

“Some methods will take existing patient data, where the diagnosis is known, and use that to try to improve the efficiency and accuracy of clinical medicine, reflecting the tasks and problems that we already use,” he says.

“However, unsupervised learning can be very useful for knowledge generation and scientific discovery, because deep learning can deal with multiple data types at enormous scale with great accuracy, to learn more and learn more quickly than would be otherwise possible.”

Reducing healthcare disparities?

One of the known pitfalls of artificial intelligence involves its potential to magnify prejudiced assumptions and predispositions, leading some to suggest that AI could also be shorthand for “augmenting inequality”[3].

However, a range of studies have found that AI can also deliver results that eliminate certain biases, such as a 2021 study published in Nature Medicine[4]. It used a deep-learning approach to measure the severity of osteoarthritis by using knee x-rays to predict patients’ experienced pain.

The study showed that the approach dramatically reduced unexplained racial disparities in pain and proposed that algorithmic predictions could potentially redress disparities in access to treatments, such as arthroplasty.

Dr Oscar Russell, rheumatology trainee and clinical associate lecturer at the University of Adelaide, says that these interventions may play a big role in helping to improve patients’ autonomy and access to care in the future.

Dr Russell and colleagues recently investigated the link between patients’ socioeconomic status (SES) and their disease outcome in rheumatoid arthritis, finding that higher SES patients had less disability soon after their RA diagnosis, and improved at a faster rate than lower SES patients – a finding that’s consistent with international data, despite Australia’s universal healthcare system.

The study also found higher SES patients received fewer opioid and prednisolone scripts and a lower overall dose of opioids.

“How we measure SES in the rheumatology literature is extremely heterogenous, which has impacts for comparisons between studies,” he says, adding that most studies used relatively few measures, often relying instead on either a personal measure or a contextual/neighbourhood measure rather than using both.

Dr Russell says that the advent of inexpensive, effective mobile health interventions may go some way to addressing these disparities. He notes that most of these interventions don’t currently use AI, with commonly used patient apps typically medication self-management diaries and symptom monitors, such as the Arthritis Australia myRA app.

AI arthritis app for GPs

One practical application of AI in rheumatology is currently under development by an Australian team led by Perth-based rheumatologist Dr Mark Reed.

The team – which includes data scientists with a medical background – has developed an AI-based arthritis screening smartphone app to help primary care physicians identify different forms of arthritis from a short questionnaire, plus a photo of patient’s hands.

Arthritis reportedly affects around 15 per cent of Australians (although some studies suggest that arthritis is significantly under-diagnosed), and osteoarthritis is the most common type.

Both osteoarthritis and rheumatoid arthritis commonly present with signs and symptoms in the hands, but while there has been excellent progress with the treatment of inflammatory arthritis, screening remains inadequate and can result in delayed diagnosis and treatment.

In May, the team published the results of their multicentre validation study of the arthritis screening app[5]. Results were analysed from a study of 248 new patients across seven private rheumatology practices.

The app combines several algorithms to accurately identify different forms of hand arthritis, predicting rheumatoid arthritis with 85 per cent accuracy, psoriatic arthritis with 95 per cent accuracy and osteoarthritis with 77 per cent accuracy.

“We designed the app to fill an unmet need for GPs, who often rely on blood tests and x-rays to screen patients who present with arthritis affecting the hands,” Dr Reed says.

These tests can be flawed, he says, because while some patients will have elevated inflammation or serology markers, the normal or negative results don’t exclude most inflammatory conditions.

“A GP without a lot of training or experience with particular conditions won’t really know what to do with those patients, and often they will just suggest Panadol,” he says.

The app guides the primary care physician through the clinical questions, weighting the responses based on intelligence gathered from a large dataset from real patients with these conditions.

“One algorithm analyses the smartphone photo that the GP takes of the patient’s hands. Then another algorithm combines the image assumptions with the responses to the survey to deliver a result – for example, it might tell the GP, ‘this looks like osteoarthritis’,” Dr Reed says. 

The whole process can take less than three minutes, so it’s easily done within a standard consultation, he says.

Because osteoarthritis can be managed by a GP, Dr Reed says that the app has the potential to reduce around 40 per cent of unnecessary referrals.

“We also often get rheumatoid arthritis patients who should have been referred to a rheumatologist earlier, but their blood tests were normal.”

Dr Reed says that the app’s accuracy in identifying arthritis sub-types could also be used for rheumatologists so that they can triage high-need patients more urgently.

“Some rheumatologists have waiting lists over 12 months; there’s a real shortage of specialists. This app has the potential to help us work smarter and have better triage and referral processes so that urgent cases can be properly prioritised.”

The arthritis screening app now has TGA approval, and the next challenge involves getting the app into the hands of GPs, says Dr Reed, which may involve commercialisation to ensure it has ongoing resourcing and relevance.

The importance of clinical experts

Dr McMaster says it is critical that rheumatology applications of AI make use of clinical experts, ensuring they are heavily involved in formulating hypotheses and developing and testing AI models.

“Domain experts need to play a critical and central role in all of this,” he says. He doesn’t agree that experts may potentially entrench bias. “On the contrary, there’s a far greater risk of increasing bias if you don’t understand the problem.”

He says that today’s medical students have probably learned computer science and programming at school, and an understanding of machine learning is already being introduced into the medical curriculum in many institutions.

“There’s no doubt that AI will play a central role in rheumatology in future, and it’s incumbent on all of us to learn about it,” Dr McMaster says.

“Clinicians need to be central in developing applications of AI because deep learning practitioners tend to only interact with the world of patients through a dataset – maybe it contains images and pain scores, and their outcome is to use this to predict pain scores.

“But the world of patients is not a data set, it contains real people and their lives, and their outcomes are about gaining a better quality of life. So, it’s important that when we apply AI we think much more deeply about what we are trying to achieve.”


[1] McMaster C, Bird A, Liew DFL et al. Artificial Intelligence and Deep Learning for Rheumatologists: A Primer and Review of the Literature. Arthritis Rheumatol 2022, online 20 July.

[2] McMaster C, Yang V, Sutu B et al. Temporal artery biopsy reports can be accurately classified by artificial intelligence –Arthritis Rheumatol 2021, 73 (suppl 10)

[3] Leslie D, Mazumder A, Peppin A et al. Does “AI” stand for augmenting inequality in the era of covid-19 healthcare?BMJ 2021, 372; 304

[4] Pierson E, Cutler DM, Leskovec J et al. An algorithmic approach to reducing unexplained pain disparities in underserved populations. Nat Med. 2021 27:136-140

[5] Reed M, Rampono B, Turner W et al. A multicentre validation study of a smartphone application to screen hand arthritis. BMC Musculoskelet Disord 2022, 23; 433

End of content

No more pages to load

Log In Register ×