After IBM's Watson cruised to victory on the game show Jeopardy!, the company decided the first area where the 'cognitive computing' system would be put to work was healthcare -- a field of sprawling data, intense privacy concerns, and life and death outcomes.
Since IBM began targeting healthcare as a sweet spot for AI back in 2011, the market for artificial intelligence in the industry has grown: in 2014, it was thought to be worth $600m, and is expected to reach $6bn by 2021, according to analyst Frost & Sullivan.
Venkat Rajan, global director of the Visionary Healthcare Program at Frost & Sullivan, says AI in healthcare has "moved from nascency and pilots and proof of concepts, to more early stage commercialisation, adoption, and utilisation".
The industry's interest in AI, Rajan says, has been driven both by rising costs and increasing volumes of data. "There isn't necessarily the capacity to capture and process and understand all of it. I think AI, particularly a lot of early solutions, are targeting those issues -- being able to take large volumes of data, put it through levels of processing that can allow some level of relevancy to crop up to support decision making and influence the course of care."
The aim is for AI systems to do what doctors can't always: keep up on every detail of every patient's visit to every specialist or hospital, as well as each pertinent new piece of research, disease outbreak, and public health recommendation. The system must not only digest all that information, but also factor in the patient's symptoms and then recommend a diagnosis or course of treatment that takes all those elements into account.
"Imagine there's a new patient that has never been encountered by a particular physician before. What if a system could draw a connection between this particular patient and to some previous cases dealt with by other physicians at other hospitals or other historic cases in the system. That could be made possible by an artificial intelligence system," said Eric Xing, a professor at Carnegie Mellon University who is involved in the Pittsburgh Health Data Alliance.
Along with IBM, creating such systems has caught the interest of another of tech's big names: Google. Earlier this year, the company's AI unit DeepMind established a health division, and announced it had been working with the Royal Free London NHS Trust on an app to help identify kidney patients at risk earlier. While the system has no AI components in its current iteration, the direction of travel is clear.
Risk stratification -- using pattern recognition to identify patients at risk of developing a condition, or seeing it worsen due to lifestyle, environmental, genomic, or other factors -- is another area where AI will begin to take hold in healthcare.
It's "trying to catch patients earlier in the disease state," Frost & Sullivan's Rajan says. "The patient that has type II diabetes has already had some sort of symptoms, but what [these systems] are trying to do is catch the patients with pre-diabetes and provide levels of support and intervention that can prevent or regress disease."
Similarly, AI has a role to play in what's being called precision medicine -- not just picking out a treatment according to the patient's disease, but also according to their history, circumstances, lifestyle, preferences, genetic makeup, and more. AI has a far greater ability to weigh up all these sometimes contradictory drives and choose the best suited treatment than human doctors are likely to.
And what's more, the same techniques could potentially be applied to drug development -- using AI systems to segment trial participants according to any number of factors and so identify why a particular medication may have worked better for one subset of individuals than another. That information could then be fed back to doctors, or their AI diagnostic assistants.
In future doctors and other healthcare personnel could also get AI assistance in other ways: the Pittsburgh Health Data Alliance's recent calls for projects in search of funding saw one proposal for a system that would automatically transcribe conversations between doctors and patients, extract the relevant material, and input it into hospital systems.
"When a doctor is talking with a patient in a typical clinical visit, then it's annoying for him to have to turn to the computer terminal and type the information. It's inefficient," Xing said.
Much of the focus for AI in healthcare is on the healthcare provider side, but what about patients? One suggestion is that in future AI systems' natural language processing abilities could be put to use directly in advising individuals on their health. Think of a new parent unsure if their baby's rash is just a skin condition or an early sign of meningitis, or someone with a sports injury not sure if they've sprained their ankle or ruptured a ligament.
Using a setup similar to Siri or Cortana, the individual could talk directly to an app, listing their symptoms and concerns, and be advised whether to take a couple of aspirin or get themselves to the emergency room.
While AI might have a lot of potential in healthcare, it has a lot of unique challenges. Take privacy for example. There's an extra importance placed on making sure data hygiene is paramount in healthcare (and that insurers' and healthcare providers' interest in that data is balanced). Healthcare also has a greater regulatory insight than many other industries, and rightly so -- AI could be involved in life and death decisions. "It's not like placing an advert on an iPhone, where if you make a mistake, no big deal, you do it again. With this, you have to be very careful," Xing said.
But as costs come down and technology improves, there's no doubt that both doctors and patients will see more of AI in future. But does that mean patients will be seeing less of their doctors?
Not so -- for the broad practice of healthcare and all the skills it entails, you can't beat intelligence of the flesh and blood kind. "By using AI, you can complement human beings' limitation in consuming large data. [AI systems] are very good at doing specific things, but they're less competent than a human doctor on a comprehensive basis," Xing said.