Doctor AI will see you now
How artificial intelligence could change medicine
- By: Peter Hateley
In March 2016, artificial intelligence hit the headlines when Google’s computer program AlphaGo beat the world champion of the complex board game Go.
But humans didn’t teach AlphaGo how to win. Instead, the program’s algorithm learnt how to play the game autonomously, using a branch of artificial intelligence known as machine learning. Machine learning feeds vast amount of information into algorithms that learn to spot patterns as they go along. In time, they get better and better at finding the outcome they are looking for—in this case, to win the game.
Machine learning is linked to deep learning, which uses models structured in a similar way to nerve cells in the brain. Jeremy Howard, founder of artificial intelligence company Enlitic, described it as “an algorithm that mirrors how the human brain works.” The algorithms are described as artificial neural networks.
Together, machine learning and deep learning give computing a huge potential, ranging from self driving cars to speech recognition. Cofounder of the DeepMind project behind AlphaGo, Mustafa Suleyman, states that “in software terms [the NHS] has been left behind for the last 20 years.” But what is the potential for artificial intelligence in medicine, and what risks does it bring with it?
Dr Google is often the first port of call for patients seeking medical advice, and around 1% of all Google searches are symptom related. A reliable algorithm could triage patients to self care where appropriate and relieve pressure on primary care. Your.MD, which is based in the United Kingdom, produces a symptom checking app that uses artificial intelligence to provide instant health advice. The app uses the information input by users to improve its understanding of how people describe their symptoms. The company has used medical test cases from Harvard University and the Royal College of General Practitioners to externally verify the accuracy of the app. Benchmark tests have shown the app has been 85% accurate in diagnosing the most common conditions and 60% accurate in diagnosing a further 500 conditions.
Another area seeing fast development in artificial intelligence is medical imaging. The ability of artificial intelligence to spot patterns from visual data is being applied to help doctors interpret a broad range of medical imaging, from computed tomography scans to funduscopy. In a recent test, three radiologists working together were outperformed by an artificial intelligence program in detecting cancers on chest imaging.
Ben Glocker, lecturer in the department of computing at Imperial College London, says: “We trained a machine that can search through multiple scans to locate, and flag up, the patterns that might indicate cancer.”
Artificial intelligence might improve diagnostic accuracy and speed, but it could go further than that, Glocker says. “Not only can intelligent machines help us routinely find the patterns hidden in complex scans, they might actually allow us to learn something new. To reveal insights that will allow us to better understand the most complex diseases.”
However, Tom Turmezei, radiology fellow at the Royal National Orthopaedic Hospital, says that although artificial intelligence has a role in making sense of large datasets found in areas like radiology, there are questions over how it will be used in imaging in the near future. He points to “clear limitations for scenarios that require skilled physical intervention, such as ultrasound and interventional radiology.”
Turmezei says that “the ability [of artificial intelligence] to search for and recognise patterns across a large volume of imaging data could yield unexpected results.” He describes the likely introduction of artificial intelligence as “a net rather than the sharp end of a stick,” providing diagnostic assistance rather than replacing clinicians. For example, artificial intelligence could flag suspicious features of an image set to a radiologist, who would then interpret them.
Google’s DeepMind project has also explored the potential of artificial intelligence in imaging. In July 2016 the company announced a research partnership with Moorfield’s Eye Hospital in London. They will use artificial intelligence to help detect common sight threatening conditions, such as age related macular degeneration and diabetic retinopathy, in more than one million anonymised scans.
Pearse Keane, a consultant ophthalmologist who set up the project with DeepMind, says: “The idea is deep learning computers would be able to diagnose common causes of retinal disease and monitor some of these conditions . . . . Early diagnoses allow us to intervene early, making it more likely to save someone’s sight.” The research will focus on advanced retinal scanning, such as optical coherence tomography. The project might help identify and prioritise patients who need early treatment.
Identifying deteriorating patients
Another application of artificial intelligence is to identify critically ill patients in hospital whose condition is deteriorating. The National Institute for Health and Care Excellence already recommends “track and trigger” systems to measure and act on patient deterioration, with many of them available on mobile devices. Artificial intelligence could take this a step further.
David Clifton, associate professor of engineering science at Oxford University, says: “A particular advantage of machine learning methods is that they can perform fusion of multiple heterogeneous data sources,” such as vital signs and lab tests, which most computers would find challenging. The ability of artificial intelligence to synthesise and analyse vast amounts of data means programs might spot patterns indicative of deterioration at an earlier stage than doctors could do.
DeepMind is also developing an app called Streams with clinicians from the Royal Free Hospital in London. It is designed to detect patients at risk of developing acute kidney injury using an algorithm based on patients’ medical histories and inpatient data. Acute kidney injury is a common cause of death in hospital patients, associated with up to 100 000 hospital deaths a year, and this innovation could help save many lives through early intervention.
Artificial intelligence is being applied in many other ways, from cancer drug research to personalised insulin injections. Machine learning requires vast amounts of data, but are there problems in using such quantities of patient information?
The data problem
DeepMind’s Streams project attracted controversy because it entailed the transfer of millions of identifiable patient records without the explicit consent of the patients concerned. The agreement was approved by the Royal Free Hospital’s data guardian and is just one of hundreds of patient data sharing agreements the NHS has with third party organisations. However, this episode highlighted the lack of public awareness as to how patient data are being shared. Patient data sharing has recently been reviewed by national data guardian Fiona Caldicott. She concluded that “people must feel able to discuss sensitive matters with a doctor . . . It is becoming ever more important that people understand when and how information is shared, how privacy is protected, and how sharing information benefits them and others.” With data driven medicine growing, and more patient data being shared as a result, these issues will become ever more important.
Neil Lawrence, professor of machine learning at Sheffield University, says the importance of regulating this area is a challenge for big data projects. “Any entity that controls large amounts of patient data will have substantial control over health recommendations across the whole NHS . . . Just as the production of pharmaceuticals is carefully regulated, so would the nature of digital innovations [need to be regulated].”
Beyond data, there are questions around where responsibility lies for artificial intelligence performance. The fatal crash of Tesla’s automated car in June 2016 showed that artificial intelligence technologies aren’t infallible. Turmezei emphasises the importance of accountability. “We would want any artificial intelligence to be held to the same standards as any human involvement in the care of a patient. Just who will be responsible for this is an interesting question,” he said.
Doctors might be suspicious of diagnostics from artificial intelligence machines that they don’t understand. Glocker says: “Understandably, doctors are reluctant to use the advice of any machine that behaves like a black box.”
Forecasts by the World Economic Forum suggest losses of 5.1 million jobs by 2020 because of artificial intelligence technologies. This seems unlikely for medics—a 2013 study by Frey and Osborne gave physicians and surgeons a probability of 0.0042 (0.42%) for being replaced by artificial intelligence in the future. The combination of skills, knowledge, and empathy might make it harder to replace doctors entirely. But will artificial intelligence be making decisions any time soon? Glocker thinks not. “I believe for a very long time we will and should have a human expert making the final decision, but with the best support we can provide using the power of artificial intelligence to extract clinically useful information from medical data,” he says.
Turmezei believes it’s unlikely that artificial intelligence will deliver a new clinical radiological sign but says the technology will be able to detect “meaningful patterns in quantitative data beyond the capability of human observation.”
For the foreseeable future, it seems that artificial intelligence will assist clinicians in providing more streamlined and cost effective patient care, rather than replace them.
Box 1: Jargon explained
Algorithm—Basis for computer processing using a set of logical steps to complete a specific task
Machine learning—Algorithms that adjust their outcome according to the input data by learning from data
Artificial neural network—Large neurone-like network of units that interact to process information. The model can adjust by putting different weights on different pathways in a way similar to a human learning by example. This enables more complex and accurate learning
University of Nottingham
Competing interests: None declared.
Provenance and peer review: Commissioned; not externally peer reviewed.
- News BBC. Google’s AI wins final Go challenge. March 2016. www.bbc.co.uk/news/technology-35810133.
- Howard J. The wonderful and terrifying implication of computers that can learn. TEDxBrussels. December 2014. www.ted.com/talks/jeremy_howard_the_wonderful_and_terrifying_implications_of_computers_that_can_learn?language=en .
- Armstrong S. The computer will assess you now. BMJ 2016;355:i5680. http://www.bmj.com/content/355/bmj.i5680. doi:10.1136/bmj.i5680 pmid:27777226.
- Pinchin V. Google Blog. I’m feeling yucky ☹ searching for symptoms on Google. June 2016. https://googleblog.blogspot.co.uk/2016/06/im-feeling-yucky-searching-for-symptoms.html .
- Carr-Brown J, Berlucchi M. Pre-primary care: an untapped global health opportunity. 2016. www.your.md/report .
- Automation and anxiety: will smarter machines cause mass unemployment? Economist 2016. www.economist.com/news/special-report/21700758-will-smarter-machines-cause-mass-unemployment-automation-and-anxiety .
- Moorfields Eye Hospital. Moorfields announces research partnership. 2016. www.moorfields.nhs.uk/news/moorfields-announces-research-partnership .
- National Institute for Health and Care Excellence. Acutely ill adults in hospital: recognising and responding to deterioration. Clinical guideline CG50. 2007. www.nice.org.uk/guidance/CG50/chapter/1-Guidance .
- England NHS. Algorithm for detecting acute kidney injury (AKI) based on serum creatinine changes with time. 2014. https://www.england.nhs.uk/wp-content/uploads/2014/06/psa-aki-alg.pdf .
- Deep Mind. Streams. 2016. https://deepmind.com/applied/deepmind-health/streams/
- Revealed: Google AI has access to huge haul of NHS patient data. New Scientist 2016. www.newscientist.com/article/2086454-revealed-google-ai-has-access-to-huge-haul-of-nhs-patient-data/.
- National Data Guardian for Health and Care. Review of data security, consent and opt-outs. June 2016. www.gov.uk/government/uploads/system/uploads/attachment_data/file/535024/data-security-review.PDF .
- Cann O. Five million jobs by 2020: the real challenge of the fourth industrial revolution. 2016. www.weforum.org/press/2016/01/five-million-jobs-by-2020-the-real-challenge-of-the-fourth-industrial-revolution/ .
- Frey CB, Osborne M. The future of employment. September 2013.www.oxfordmartin.ox.ac.uk/downloads/academic/future-of-employment.pdf.
- Published: 03 January 2017
- DOI: 10.1136/sbmj.i6528