The article looks at the use of artificial intelligence (AI) medical diagnostic applications (apps) used by the British National Health Service, in order to reduce the error disease diagnosis and drug prescription. Particular focus is given into the impacts of errors in given by physicians in diagnosis of disease in patients contributing to the unnecessary deaths. It also emphasizes that despite the use of AI, physicians are essential to create empathic relationship with its patients.
Artificial intelligence will soon be a standard part of your medical care — if it isn’t already. Can you trust it, asks Kayt Sukel
THE doctor’s eyes flit from your face to her notes. “How long would you say that’s been going on?” You think back: a few weeks, maybe longer? She marks it down. “Is it worse at certain times of day?” Tough to say — it comes and goes. She asks more questions before prodding you, listening to your heart, shining a light in your eyes. Minutes later, you have a diagnosis and a prescription. Only later do you remember that fall you had last month — should you have mentioned it? Oops.
One in 10 medical diagnoses is wrong, according to the US Institute of Medicine. In primary care, one in 20 patients will get a wrong diagnosis. Such errors contribute to as many as 80,000 unnecessary deaths each year in the US alone.
These are worrying figures, driven by the complex nature of diagnosis, which can encompass incomplete information from patients, missed hand-offs between care providers, biases that cloud doctors’ judgement, overworked staff, overbooked systems, and more. The process is riddled with opportunities for human error. This is why many want to use the constant and unflappable power of artificial intelligence to achieve more accurate diagnosis, prompt care and greater efficiency.
AI-driven diagnostic apps are already available. And it’s not just Silicon Valley types swapping clinic visits for diagnosis via smartphone. The UK National Health Service (NHS) is trialling an AI-assisted app to see if it performs better than the existing telephone triage line. In the US and mainland Europe, health insurers and national healthcare providers are hopeful AI-based medical apps will improve care. But is the hype around medical AI all it’s cracked up to be? Would you trust your care to a robot?
For decades, researchers have been honing artificial intelligence, including deep-learning algorithms, which are designed to learn without being fed rules or constraints (see “A glossary of AI speak”, page 38). “These would take in hundreds or even thousands of symptoms and then would learn to diagnose various diseases,” says Pedro Domingos, a computer scientist at the University of Washington and author of The Master Algorithm: How the quest for the ultimate learning machine will remake our world. “By training these systems with the data from a medical database of patient records for, say, diabetes or lung cancer, or any other condition, you can push a button and literally get something that will diagnose things more accurately than human doctors can.”
That’s not just hype. When Sebastian Thrun and his team at Stanford University in California trained a deep-learning neural network using more than 100,000 images of skin problems, ranging from cancer to insect bites, then tested it on 14,000 new images, the system correctly diagnosed melanomas more often than seasoned dermatologists. Deep-learning networks have also outperformed doctors at diagnosing diabetic retinopathy, a complication of diabetes that damages blood vessels in the eye. Other AI tools can identify cancers from CAT scans or MRIs, or even predict from data about general health which people may have a heart attack.
But should we trust their successes? What are these systems seeing that highly trained doctors aren’t? It’s not a question that can always be answered. While some deep-learning tools are designed to spit out the rules they come up with, Thrun’s, for instance, was a “black box”: it’s unknown what features it homed in on.
That makes some nervous, with reason. Consider the experience of Joshua Denny, a medical informatics specialist at Vanderbilt University in Tennessee. He recently developed a machine-learning tool to identify cases of colon cancer from patients’ electronic health records, but soon learned that it was latching on to the wrong information. “It was performing excellently,” he says. Unfortunately, it was picking up on the fact that all of the patients with confirmed cases had been sent to a particular clinic, not clues from their actual medical data. “There’s always the risk that a black box model can learn features that you won’t expect — and won’t be stable over time,” he says.
While acknowledging potential pitfalls, Thrun is circumspect about the nature of the black box approach. “If your doctor looks at your skin and says, ‘I think this is a melanoma,’ you aren’t going to stop him and say, ‘What are the rules you are using to determine this?'” he says. “No, you are going to have a biopsy and then, most likely, get treatment. We shouldn’t distrust these rules just because we can’t say exactly what they are.”
Still, Thrun concedes that deep-learning tools are only as good as the data they are trained on. Thanks to the rise of electronic medical records, we finally have big enough data sets to do this training, but there are major logistical hurdles to overcome. The wide range of healthcare IT systems can mean that their records vary just enough to skew any algorithm trying to process them. That means up to 90 per cent of the effort in designing these AI tools is spent simply cleaning up the data, Thrun says. Kyung Sung, a radiologist at the University of California, Los Angeles, agrees. It took his team more than five years to clean up a set of prostate cancer images for his AI, which aims to better identify aggressive tumours. “Unfortunately, the fact that we now have all these images available doesn’t always mean that they will be useful,” Sung says.
It’s even possible that an emerging or evolving disease might skew the results, which is something developers have to look out for. For example, the creators of diagnostic app Ada trained it on vast troves of medical files, and it now refines its results with data from users. To avoid the AI picking up on the wrong things and warping the outcomes, human-supervised and unsupervised learning are both used to fine-tune the algorithms, says Claire Novorol, the company’s chief medical officer. “We use multiple experienced doctors as well as other technical feedback loops,” she says.
Humans keep the AI in check, but there will be times when they shouldn’t. “If there are presentations not currently known by experts or in the literature, humans may not be the best ones to catch that,” says Novorol. “Those trends will be patterned in the data and, ultimately, patient outcomes — and the algorithms can help us identify those patterns and use them in a predictive way.”
Today, deep-learning diagnostic tools are not used in hospitals except in research studies, but many think they will be within five years. By then, they will be able to do far more than diagnosis, says Eric Horvitz, a medical AI researcher at Microsoft. “The hard part is managing diseases, figuring out therapies over time and tracking progress,” Horvitz says. New, more detailed algorithms should help doctors better understand how conditions progress. “Diabetes, arthritis, hypertension, asthma and other chronic diseases — these are the expensive, challenging cases. These are where most of the healthcare costs come in. Machine learning may offer us new opportunities to better manage them.”
Horvitz is not alone in his optimism. Valentin Tablan, principal scientist at Ieso Digital Health, sees potential for AI to revolutionise mental healthcare. Ieso provides cognitive behavioural therapy online for NHS patients and has treated more than 10,000 people to date, keeping digital records of every exchange. The company wants to mine that immense data set to help understand what really works. “Machines are very good at helping find the elements that are very important and can really help patients get better,” Tablan says.
With some studies suggesting that patients may be more open with therapists when talking to them via a computer screen, is it time to consider removing the human altogether? Tablan scoffs at the idea. “AI doesn’t have the capabilities to work at that kind of level yet. But by building models based on these data sets, we can create a tool for human therapists to use that makes them superhuman therapists.”
That’s a recurring theme: the rise of the superhuman doctor. By equipping medical professionals with enhanced abilities, AI is poised to change the very delivery of healthcare, says Isaac Kohane, head of biomedical informatics at Harvard Medical School. At present, doctors have to manage mounds of paperwork and digital form-filling while trying to stay on top of the emerging research to keep their knowledge current. If AI could ease some of this burden, that would free them to focus more time on patients, to take detailed histories, to listen.
Ultimately, it may even reshape what it means to train as a doctor. Denny says medical education will need to include data science and may shift away from rote learning to focus on problem-solving, critical thinking and how to best deal with the probabilistic outcomes that so many AI systems produce.
As medical AI matures, Thrun believes it will replace many roles in dermatology, radiology and pathology — those that mainly involve repetitively reviewing images. At the same time, there will be growth in other areas, such as specialised surgery.
It could also change what it means to be a family doctor, says Kohane. “They will be able to offer their patients specialty care, like imaging and dermatology procedures right in their office, with expert-level performance — and then refer the patient to a specialist if and only if a truly actionable finding comes up.” That could mean a more holistic approach, and not having to split care between half a dozen doctors. “That would be a great thing both for the doctor and the patient.”
There are a few significant obstacles to jump first. To start: how to provide the massive data sets that AI systems need, while protecting patient privacy. The advent of electronic medical records has also ushered in stringent regulations, such as the HITECH Act in the US and the Data Protection Act in the UK. Last year, New Scientist discovered that the NHS had shared patient data with Google DeepMind, a deal the UK Information Commissioner’s Office just found “failed to comply with data protection law”. The irony is, for medical AI to truly take off, even more rapid and wider sharing of data may be necessary. That might require new legislation. “Current laws don’t really cover the kind of sharing scenarios we need to make these systems work,” Denny says.
Domingos agrees that the legal framework will need to change, and stresses that any policies must require informed consent. But he also argues that sharing your health data should be seen as a civic duty, and that only those who opt in should reap any benefits. “If someone won’t allow their data to be used, then they shouldn’t have access to the better treatments that result,” he says.
Even as the debate over privacy flares up, there’s still the matter of liability. Malpractice laws are complex and vary from place to place, so it’s unclear how they might need to change to accommodate AI. Kohane isn’t worried, though. He points out that doctors already use machines to make a diagnosis — software that helps them identify tumours in MRI scans or abnormalities in echocardiograms, for example. “If a doctor is in the loop, the legal and ethical stuff is not going to be that challenging,” he says; ultimately, it’s the doctor’s responsibility. If AI and doctor disagree, a supervising physician or committee could break the tie.
Standalone AI systems would require further consideration, however. “That is indeed terra incognita and something we’ll have to figure out as we go along,” Kohane says. Liability may ultimately switch from the physician to the manufacturer, as with self-driving cars. “Volvo has said they will assume liability for their self-driving cars. For the kind of AI machines you might find in your corner pharmacy, the company that made that machine is going to have to assume liability for its range of parameters,” says Denny. As for what happens if that AI gets it wrong, “it’s something we need to really think carefully about”, he says. “To my knowledge, the malpractice industry hasn’t yet thought about this kind of thing. But it’s time that we do.”
Once these thorny issues have been worked through, the question is whether standalone AIs will ultimately replace doctors. Not likely, says Denny. By streamlining diagnosis, they will make it easier to access credible medical advice no matter where you live and will assist with a lot of routine care. “These systems will allow physicians to reduce their mental load, to pay more attention to each patient, to prioritise which patients need critical care right now — to be more efficient overall,” says Denny. “It’s going to be a win for everyone.”
Doctors won’t be cut out of the picture, because their empathic relationship with the patient is an essential part of care, says Vimla Patel, a cognitive psychologist and specialist in biomedical informatics at the New York Academy of Medicine. AI can augment clinicians’ abilities, but can’t do all the heavy lifting. “When things get complex, and medicine often is complex, you need human reasoning to make decisions,” she says. “Computers, no matter how sophisticated, cannot replace that.”
Artificial intelligence Applying computers to tasks that normally require human-level intelligence, like reasoning, decision-making, problem-solving and learning
Big data The huge data sets that can be analysed by computers and algorithms to reveal patterns, trends and associations
Machine learning The capacity of an algorithm to learn from new information and modify its processing as a result, without being explicitly programmed to do so
Neural network An algorithm used in deep learning that imitates the activity of layers of neurons in the brain, filtering data through tiers of virtual brain cells
Deep learning The “black box” of AI. Unsupervised neural networks that create their own processing constraints as they learn from vast troves of training data
An AI could spot the significant picture much faster than a doctor
~~~~~~~~
By Kayt Sukel
Kayt Sukel is a writer based in Houston, Texas