ARTIFICIALLY intelligent doctors are here. Thousands of people in the US and Europe have already been screened by an AI system for detecting diabetes-related blindness without the involvement of a human doctor. The system was approved last year after it .
More AI tests will kick off in the next few years. They look set to improve the diagnosis of many conditions, from breast and to broken wrists and glaucoma. Any hospital that can afford the necessary equipment will soon be able to offer the same standard of diagnosis.
This is the vision, at least. Yet if we rush to adopt such systems prematurely, they could prove harmful. 鈥淚鈥檓 bullish about the ability of AI to do good,鈥� says Amol Navathe at the University of Pennsylvania. 鈥淏ut it鈥檚 harder than people think.鈥�
Advertisement
Take IBM Watson, the AI system famous for winning the Jeopardy TV quiz. It was supposed to revolutionise healthcare by using its natural language processing skills to analyse vast amounts of medical literature to provide more accurate diagnoses and recommend better treatments.
But it has been claimed that the cancer system, for Oncology, sometimes made incorrect and unsafe recommendations 鈥� which weren鈥檛 followed 鈥� and that are abandoning the system after it didn鈥檛 live up to expectations. 鈥淚n my view, it鈥檚 a failure,鈥� says cardiologist Eric Topol, author of Deep Medicine, a book on AI in healthcare.
IBM has disputed these claims. 鈥淐an Watson help oncologists make better decisions for their patients? Repeatedly, the answer has proven to be a resounding 鈥榶es鈥�,鈥� .
But Navathe says he and others have found that you can鈥檛 just feed a whole lot of data to an AI and expect it to produce brilliant insights, because of inherent biases in data sets. 鈥淭he data that they are seeing is filtered through the eyes and judgement of humans,鈥� he says.
For instance, an AI fed hospital test data might seem good at predicting what disorder a patient has long before the results come back. But because doctors order tests based on what they suspect is wrong with a person, what the system might really be doing is using its knowledge of which tests have been ordered to work out what doctors suspect the problem might be, which isn鈥檛 helpful.
Some of the pitfalls are even more subtle. Navathe has worked on AI systems for spotting sepsis, a sometimes deadly immune response to infection. The aim is to predict early who might die to give more time to save lives.
The trouble is, those who die often have little chance of survival because of other health issues. So even if an AI can identify these people, if doctors follow its advice, they might end up aggressively treating people who should instead be getting palliative care to ease their final days.
These issues arise because the most popular form of AI, deep learning, is basically a form of pattern recognition. Deep-learning systems are a black box, so there is no easy way to tell if the pattern they pick up on is a result of a bias in the data.
Some groups are instead developing AI systems built around a model of cause and effect. These causal models are designed to be transparent, to explain their decisions.
For now, though, most groups are applying machine learning to relatively straightforward, objective forms of data such as MRI, CT and retinal scans.
The US Food and Drug Administration has already approved about a dozen AI algorithms, such as the OsteoDetect system for spotting broken wrists in X-rays and the DreaMed Advisor Pro, which uses glucose monitoring to recommend dosages for people with insulin pumps.
There is much more in the pipeline. IDx, the firm that developed the test for , is also working on diagnosing macular degeneration, glaucoma, Alzheimer鈥檚 disease, cardiovascular disease and stroke risk from retina scans, for instance.
鈥淔or macular degeneration and glaucoma, we expect to start clinical trials by 2020,鈥� says IDx spokesperson Laura Shoemaker.
DeepMind, another AI company, has demonstrated a similar system for detecting problems on retinal scans.
Yet even with these relatively simple applications of machine learning, there are issues. First, if an AI is trained on one make of imaging equipment, it will work only with those machines. In fact, even upgrading imaging-machine software can confound AIs.
Ideally, systems should be generalisable. What鈥檚 more, if they are trained on scans of only people of European descent, say, they are unlikely to work for other people.
鈥淚鈥檓 bullish about the ability of artificial intelligence to do good, but it鈥檚 harder than people think鈥�
鈥淭his is an extremely important point,鈥� says Constance Lehman at Harvard Medical School, whose team has created a system that outperforms doctors at analysing mammograms. It is a problem in medicine generally, not just in AI, she says.
Perhaps the biggest issue is ensuring that people really do benefit. While we tend to assume that the more tests done the better, this isn鈥檛 the case.
When it comes to breast cancer screening, for instance, a 2013 study concluded that saved, 10 women had unnecessary treatment and 200 suffered years of needless stress. Such findings have led some to argue that mass mammogram screening should be abandoned. 鈥淲e need to be careful about overdiagnosis,鈥� says Lehman.
Similar problems might come from personal tech. Take Apple鈥檚 latest watch, with an ECG function that can identify abnormal heart rhythms with the help of an AI algorithm. Many people have irregular heart rhythms without symptoms. These people might be put on unnecessary drugs or even have needless surgery.
Navathe thinks regulators should demand evidence that such tests really benefit people. Earlier this year, he and his colleagues to be assessed on the basis of criteria such as extending people鈥檚 lives. Where systems involve adaptive algorithms that learn on the go, there need to be regular audits.
He is concerned that regulators are applying lower standards to algorithms than to drugs or devices. 鈥淏ecause they are not invasive, they seem lower risk.鈥� But, he says, if doctors are basing treatments on AIs, the same standards need to be applied.
Even those pointing out the problems with AI, like Navathe, are optimistic about the long-term prospects, though (see 鈥�Medical monitors鈥�). AI really does look set to make medicine better, more equal and maybe even more human.
Topol thinks AI could free doctors from being glorified 鈥渄ata clerks鈥�, enabling them to spend more time with the people who need their help. Many studies show that outcomes are better when patients have a relationship with their doctors, he says.
But it will only happen if doctors stand up against the managers and pen-pushers, he says. 鈥淚t鈥檚 going to take unprecedented action.鈥�
Medical monitors
The AI-based medical systems being rolled out are essentially ways of diagnosing diseases faster and more accurately. As such, they fit neatly into the existing model of healthcare.
But some people think AIs will end up transforming the nature of healthcare. Instead of going to see a doctor only when we are sick, our health will be constantly assessed by AIs, based on data from devices such as smart watches along with your genome sequence. These AIs will alert us to potential problems before we are aware of them ourselves.
At least this is the vision of companies like Babylon 午夜福利1000集合, maker of an app for connecting people to doctors via an AI triage system. 鈥淭he system will become part and parcel of your life, constantly monitoring your health,鈥� says Saurabh Johri, the firm鈥檚 chief science officer. And with millions of people using systems, disease outbreaks could also be detected and tracked in real time, he says.
But this vision is far from becoming reality. Babylon鈥檚 AI is essentially a chatbot designed to help assess people鈥檚 symptoms, and so far it .
In fact, not everyone is convinced these sorts of systems can even be described as AI. 鈥淚 would view them as productivity tools,鈥� says Amol Navathe at the University of Pennsylvania.
