Have you asked Alexa to play some country music, call you an Uber, or turn up the heat? Does your new car inform you when you've departed from your lane, slow you down on the highway to keep you from tailgating, or parallel park for you? Did Amazon offer helpful suggestions while you were doing your holiday shopping?
Artificial intelligence (AI), the next transformative leap in technology, is quietly weaving itself into the fabric of our daily lives. The personal computer, the Internet, and the smartphone each made irreversible changes in how we live. (If you doubt this, try to count the ways you use all three in the course of your day, and try to remember, if you're old enough, what your day was like before.) AI will do the same over the next decade.
AI is formally defined as "the study and design of intelligent agents," or computer systems that perceive their environment in some manner and respond with actions to maximize their chance of success – activities generally associated with intelligent beings. Success can be parking within the lines, appropriately identifying impending equipment failure or correctly interpreting a radiology image.
AI is an umbrella term that can cover a range of capabilities: voice recognition, contextually aware interactions (for example, "people like you bought these products"), pattern identification (for example, finding evidence of overlooked medication effects by analyzing data on thousands of patients, or modeling traffic light sequences to minimize rush-hour slowdowns), or complex process automation (for example, extracting information across multiple systems and documents to create an accurate bill).
The AI ecosystem
The idea of AI has been around since the mid-1950s, initially as a sci-fi conceit and, then, as an ever more sophisticated Holy Grail for computer scientists. Only now is it starting to play a noticeable role in daily life, however. Why? Because we finally have an "ecosystem" that can support it. All new technologies need an ecosystem to give them traction. Cars needed roads, gas stations, traffic laws, and people who wanted to go places. Mobile devices needed small but powerful chips that wouldn't overheat, as well as high-speed wireless networks, location awareness, long-lasting rechargeable batteries, and an abundance of easily accessible apps.
AI needed, and now has:
• Processing power. Computers keep getting more powerful and efficient, and programming techniques like deep learning are enabling them to handle information the way brains do. Millions of computers can work together to tackle the most complex computations. Cloud services allow any individual or enterprise to tap extraordinary amounts of storage and processing capability on demand.
• Environment-awareness technologies. Highly miniaturized and specialized sensors and the "Internet of Things" can generate data for, and execute instructions from, AI-enabled systems. Advances in packaging can place them almost anywhere, from outer space to inside the body.
• Staggering amounts of very diverse data. From the human genome to Web browser cookies to cellphone traffic to CT and MRI scans, we have more information than we can understand unaided. We currently generate 2.5 quintillion bytes per day – and that number is not going to shrink.
• A demand for insight. The more we realize we can know, the more urgently we need to know it. Is our car going to break down? Can our network be hacked? Have we ordered enough Fingerlings to make it through Christmas? Is this a stolen credit card? Am I going to enjoy this movie? Or my upcoming date with this new person? As a society, we have moved from being slightly suspicious of AI to being impatient that it can't do more, and more quickly.
• Maturity of business models. Across all industries, organizations that provide products and services are beginning to understand how to leverage AI to improve their offerings and make them more competitive. AI adds value in all kinds of ways, from equipment that alerts users of impending failure, to shopping recommendations for a spouse's upcoming birthday.
AI will transform health care, too
As a field drowning in information with a life-or-death need to understand it, health care is ripe to benefit from AI. Certain forms of AI have been built into electronic health records and other clinical computing systems for years, under less sexy names like "decision support." Predictive algorithms can identify patients at high risk of readmission, or give early warning of impending sepsis. Care guidelines and protocols trigger reminders for clinicians to order certain tests or schedule follow-up appointments.
We can expect AI's abilities to develop rapidly in the following realms:
• Extraction of data and structures. The ability to "understand" natural language, images and video will allow AI-enabled systems to extract quality measures from clinical data, make treatment suggestions, and auto-reconcile inconsistencies, gaps and errors in clinical data. Machines can read the bits from digital imaging directly, without the need to generate an image that a radiologist could read. We are in the early stages of these capabilities; computers still make significant errors when trying to distinguish signal from noise, or an organ from a tumor. But AI could potentially catch details that elude human readers. AI can look for patterns in data and identify whether a medication is harming a patient, which patients might be candidates for a beneficial off-label use, and how social factors may be affecting a patient's health and the effectiveness of a care plan
• Cognitive interaction. AI-enabled electronic health records (EHRs) will understand context. They will tailor the presentation of patient data and care recommendations, based on an analysis of the patient’s conditions, the caregiver’s preferences, the patient’s preferences, the evidence, and insurance requirements. This tailoring eases the cognitive burden on the clinician. Contextual awareness will also improve the systems that patients use directly, so that they offer the most appropriate information in an easily understandable form.
• Operational process models. AI-enabled systems will take population health planning to the next level, tweaking care plans on the fly based on individual and community changes – events such as the death of a patient's spouse that might trigger a depression, or a jump in pollen count that increases the risk of attacks for all the neighborhood's asthma patients. AI-enabled systems will monitor the flow of patients through the hospital and streamline activity; for example, delaying the transportation of a patient to radiology until the phlebotomist, one floor below, can complete his rounds.
• Clinical models. Predictive analytics will increasingly be used across a range of situations – readmissions, transitions of care, financial clearance, and medication compliance – and AI will allow analytics to learn and improve. Best practices will increasingly be defined through machine learning (based on EHR, claims, and device data), which will identify what works and what doesn’t. Constant monitoring of the data will keep the best practices current.
Health care AI will also show up in more mundane forms: for example, in medical equipment and devices that can monitor their own health and let technicians know when they need maintenance, are about to fail, or are improperly configured.
But buckle up and prepare for a hype storm. As a recent article in the Harvard Business Review observes, grandiose AI promises may dribble away into disappointment. IBM's Watson was supposed to be the oncologist's best friend, diagnosing cancers and choosing the best treatment path from hundreds of options. More than $60 million and four years later, test site M.D. Anderson Cancer Center put the project on hold for lack of satisfactory progress and tangible results, though its IT group is successfully using AI capabilities to do less ambitious jobs like making tailored hotel and restaurant recommendations for patients' families.
What you need to know now
AI will raise some urgent issues for health care providers. They include:
• Finding and retaining the staff needed to support these systems, amid a general shortage of data scientists.
• Clarifying liability: who's responsible if a driverless car hits someone, and who's responsible when medical AI malfunctions?
• Determining when AI-based systems are “medical devices” – a topic that will continue to confront the FDA.
• Securing the systems from hackers and malware, and making sure that their self-maintenance functions are reliable.
Fortunately, health care providers won't need to shop for, or evaluate, AI as such. It will be a component of technology they are shopping for anyway. Salespeople may sling AI jargon like Singular Value Decomposition, Restricted Boltzmann Machines and Random Forests – but don't worry. Customers don't need to understand how AI works, any more than they need to understand how a cell tower works in order to take a call in their car.
However, as with any product, be prepared to ask some questions of your application, analytics, and device suppliers. Here are a few to start with:
• How are you applying AI in your products? What can your products do, as a result, that they couldn't do before, and why do I want that? (Those same questions can be used with any vendor, whether it's MRI or building control systems.)
• How have you tested these capabilities? What were the results? Do we have to change our processes or our organization to achieve comparable results and, if so, how?
• Where do you get the rules and logic that your AI is based on? How do you keep your logic current? How do you incorporate new rules (especially to keep up with the torrent of new medical knowledge being published daily)? How often do you update?
• Is there anything special that I need to teach my staff? Will the application interface change routinely? Will the product behave in ways that they don't expect? Will the product change as it "learns"?
• How am I protected in the event that your AI makes a mistake? What are the liability provisions in your contract?
AI's impact, though it will explode over the next decade or so, is already starting to be measurable. Amazon has noted that its Echo users spend an average of $1,700 per year while its regular customers spend $1,000 per year. In the health care realm, risk predictive algorithms have reduced unnecessary hospitalizations, and sepsis detection algorithms have saved lives. It is not too early to be looking at ways that AI can help you accomplish your clinical, operational, and financial goals.
John Glaser, Ph.D., is senior vice president of population health with Cerner in Kansas City, Mo. He is also a regular contributor to AHA Today.