Final Frontier of AI in Clinical Diagnostics Getting Closer

Liam Davenport

November 11, 2019

GLASGOW — The promise of artificial intelligence (AI) to improve clinical decision-making and diagnose cancers earlier is on the horizon, with some successes already seen, but there remain a host of unanswered questions that have social and ethical implications, say experts.

During a dedicated session at the National Cancer Research Institute (NCRI) Cancer Conference 2019, the concept of AI and its application to diagnostics was explored from many angles, with real-world examples revealing the future for clinical care, and one study showing that the technology significantly improved cancer detection rates.

Still a Role for Humans

The optimism was nevertheless leavened with warnings over machine learning systems being only ever as good as the data they are built on, and the unknowable nature of automated algorithm development, which could have unforeseen consequences if, as is already the case, relatively untested systems are used in clinical situations.

Prof Nicola Strickland, a consultant radiologist at Imperial College Healthcare NHS Trust, London, who discussed the implications of AI in radiology, reassured the audience, however.

She said that "nobody is actually suggesting that AI or deep leaning or whatever you want to call it, is actually going to take over from humans", but instead it will form a "symbiosis" with clinicians to help them in focusing on where they can make the greatest difference.

She added: "It should free me up as a radiologist to interact more with my patients.

"What do patients want to know? They want an interpretation of their imaging result," but, she said she can sometimes be too busy reporting and doing biopsies to see patients.

Dr Paul Brennan, senior clinical lecturer and honorary consultant neurosurgeon at the University of Edinburgh, who also spoke during the session, suggested that it is possible to overthink the ideas around AI.

"You're sometimes sat there thinking: Well, how do I know what I know. What evidence do I have for what I know? And you can get into all sorts of slightly Kafkaesque circles."

For him, AI should focus on creating solutions and then testing and validating them in a real-world population before starting again from the beginning.

"The real challenge to all of us is to be agile and, if we can do that and we can adapt all these different technologies, then hopefully that's what will drag the field forward," he said.

Tackling Delayed Diagnosis

Opening the session, Dr Brendan Delaney, professor of medical informatics and decision-making at Imperial College London, said that the issue that is driving interest in AI is that of diagnostic error.

The greatest error, accounting for 29% of malpractice claims in one US study, is delayed diagnosis, which has the gravest consequences in conditions such as cancer, neurological disorders, and vascular disease.

In addition, he said, errors may include clinicians not using symptom checkers or disease-specific scores in electronic health record systems.

While AI could be the solution, Prof Delaney underlined it is not without its own potential issues, the most fundamental of which is that a machine learning system trained on biased data leads to biased models.

It is also often not possible to know how the machine is learning, and so the process becomes something of a "black box".

Even if these issues are overcome, it is crucial that AI systems are integrated into care processes, rather than replacing them, although that does not mean that they would be used.

Together, Prof Delaney said, these create a whole host of legal, ethical, and social implications.

Learning Health System

A better approach, he explained, is to develop a Learning Health System (LHS) in which data, knowledge, and performance feed into the process in a continuous loop that is curatable, transparent, traceable, and scalable, even while it learns and adapts.

This, however, requires infrastructure, both physical and in terms of data standards; clearly-defined outcomes and potentials for bias, and methods to control them, as well as defined expectations as to how the system would be used to support human judgement.

Prof Delaney gave several examples that underlined the way in which an LHS could offer diagnostic support while, via electronic health records, feed data into an analytical system that combines it with other sources of information to improve the future quality of the diagnostic support given.

In the second presentation, Prof Strickland looked at the potential for AI in imaging to improve early cancer diagnosis.

She said that this would not only help to treat and cure the disease, and so help to reduce suffering, but also save money, through direct treatment costs and societal and economic costs.

AI could help, she said, by optimising care coordination to improve the delivery of care in a timely manner. In radiology in particular, it could be used to recognise normal, non-cancerous scans.

This would allow normal images to be "de-prioritised" to the end of the list, and "allow radiologists to concentrate on the abnormals".

Consequently, Prof Strickland said, patients with serious pathology would not have to wait weeks for a diagnosis and, in the end, the reporting of normal scans would be speeded up.

Reports could also be "prepopulated" with information that can be reliably extracted by an algorithm analysing an image, such as a patient's metastatic burden or the size of metastatic lesions.

There are also trials in the UK of AI as a "third reader" for breast screening, with "smart mapping" of suspicious areas and breast density assessments.

Risk Prediction

For Prof Strickland, the greatest use of AI could be in cancer risk prediction, taking into account changes in tumour behaviour over time resulting from tumour heterogeneity and genetic evolution.

She pointed out that biopsies are unsuited to monitoring tumour kinetics, as they can only be taken from certain parts of a tumour at any one time and may not include all metastatic sites, with the consequence that they are "inevitably skewed" and "may be significantly misleading".

On the other hand, genetic mutations are what drive tumour behaviour, and knowledge of them can help clinicians reduce over- or ineffective treatment, she said.

Combining imaging with an understanding of the genetic profile of primary tumours and their metastatic sites could allow selection of the most appropriate drug or drugs, each with their percentage likelihood of achieving a complete response.

This concept of the "virtual biopsy" is already being explored through radiomics, or the quantification of the phenotypic features of a lesion from medical imaging, and radiogenomics, which takes this concept to the next level with genetic information.

A tumour's "radiomic signature" could therefore predict a lesion's diagnosis, prognosis, and therapy response, delivering image-based precision personalised medicine.

However, Prof Strickland cautioned that this is dependent on having high-quality "ground truth" datasets that are large and well-annotated and acquired through uniform methods.

They also need to be regulated and safe before being integrated into the clinical radiological workflow.

She said that the introduction of algorithms into clinical care is already happening, "which is quite frightening", as they may not have been "properly tested in a clinical setting".

"How would I know if an algorithm developed in, say, Tel Aviv, Israel, for detecting abnormalities on a mammogram is going to work on a population of women in Taunton, in Somerset?" They have used a completely different dataset for developing, validating and testing the algorithm.

Brain Tumours

Next, Dr Brennan gave an example of how AI is already being explored clinically in the diagnosis of brain tumours.

He pointed out that, while brain tumours are not that common, "they have a significant impact on patients, and the number of years of life lost for patients with brain tumours is amongst the highest".

The most common form of the disease is glioblastoma: "It kills people quite quickly," Dr Brennan said.

He noted that the standard of care has not changed for the past 20 years.

"It's not a great position to be in, and that paucity of progress in patient outcomes and how we enhance survival is really lagging behind a lot of the other cancers."

Dr Brennan said that it's "not for want of trying", but there has been "an absolute lack of progress of data coming into the clinic".

Recognising that one area of potential improvement is in diagnosing patients earlier, he explained that systems based on symptom analysis have, however, not had much success.

He and his colleagues therefore developed an AI-led system that combined infrared spectroscopy and machine learning to analyse the spectrum of blood samples from individuals with suspected brain tumours.

As reported by Medscape News UK, the technique was found to have a sensitivity in detecting brain tumours of 81%, rising to 92% in patients diagnosed with glioblastoma.

Dr Brennan concluded that techniques such as that could have a "transformative impact" on cancer referral pathways.

In another session at the meeting, Dr Bhavagaya Bakshi, a general practitioner and co-founder of C the Signs, said that one of the main challenges in diagnosing cancer early is that on average general practitioners diagnose only around six to eight cases per year, and see a rare cancer once in their lifetime.

Cancers are also hard to diagnose, as each of the more than 200 types of cancer have their own signs, symptoms and risk factors, so the process is further hampered by limited time with patients and issues with access to tests.

To help general practitioners navigate the myriad options when assessing patients, Dr Bakshi and colleagues developed a digital clinical decision support tool that combines AI with the latest guidelines and research to cover the entire spectrum of cancer.

Crucially, the tool is adapted for each practice area by incorporating local presentation rates for each type of cancer and other regional data, so that the tool offers clinicians the most relevant information.

At all stages, the data included in the system is monitored and controlled, and the AI is not designed to undertake any independent machine learning.

To test the system, they administered it to three clinical commissioning groups in England, within which 286 clinicians from 85 practices used the tool.

This resulted in 2084 patients being assessed, with the tool used on average more than 75 times per week, with peak activity during clinical hours.

The team saw that there was a 6.40% increase in cancer detection rates in the three pilot sites during the study period, compared with an increase of 0.21% in neighbouring areas and 0.59% in England as a whole (p<0.05).

The pilot sites also experienced a 7.09% reduction in emergency cancer presentations over the study period, versus a 5.75% reduction in surrounding areas and a 4.49% reduction across England.

Dr Bakshi pointed out that, in the main, low cost tests were recommended by the tool in just under 65% of cases, with relatively few referrals or use of direct-access diagnostic tests.

She concluded that this is the "first AI-driven tool that has had a statistically significant impact on cancer detection rates".

Dr Brennan disclosed consultancy fees paid to the University of Edinburgh for work with Clin Spec Dx, Ltd. Dr Bakshi is a co-founder of C the Signs. No other potential conflicts of interest declared.

No funding declared.

NCRI Cancer Conference 2019: Artificial intelligence – does it have promise for cancer prevention and early diagnosis? Presented 4th November.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.
Post as: