This article was originally published in our Fall 2018 print issue.
From Hephaestus’s creation of mechanical men in ancient Greek myths to Arab inventor Al-Jazari’s design of the first programmable humanoid robot in the 13th century, the concept of turning inanimate objects into intelligent beings has been a long-standing fascination by humans. While the term “artificial intelligence,” or AI, was not coined until 1956, the seeds of modern AI were first planted by classical philosophers thousands of years ago.
The hype surrounding modern AI has not always been constant. Rapid phases of progress have been quickly followed by periods of reduced interest and funding called “AI winters.” Despite the deep historical roots of machine intelligence, the successive combination of the two letters “A” and “I” has the profound ability to elicit polarized responses. On one hand, we hear about artificial intelligence offering practical solutions to a variety of real world problems. Self-driving automobiles to tools like Google Translate give us hope of a future where life is even easier and more efficient. Buzzwords like “neural networks” or “predictive analytics” have given rise to multi-billion dollar investments by companies like Google, Amazon, Apple, and Facebook, which are dedicated to further researching and applying these concepts. On the other hand, many people fear a future in which computers take over the world.
Regardless of this polarization in opinion, many can agree that AI has revolutionized a multitude of industries. Healthcare is no different. The healthcare sector has been at the forefront of the development of various AI technologies, adopting more and more innovative solutions to improve prevention, diagnosis, treatment, and long-term care.
Before we choose to either jump on the AI bandwagon or abandon ship, it is important to consider the facts. AI is centered on the premise that human beings are not always the most rational creatures. Humans often lack the information and capacity needed to achieve certain favorable outcomes or “success events.” Even though reaching a “success event” is not necessarily guaranteed with machine intelligence, it is guaranteed that the machine has optimized its search for the best strategy towards a solution. In fact, a recent analysis from Accenture, a leading global professional services company, considered investments, revenue growth, and acquisitions in the artificial intelligence space. According to their findings, AI applications have the potential to create $150 billion in annual savings for the U.S. healthcare economy by 2026. The landscape of medical problems that can benefit from AI is constantly expanding.
Both classical and machine learning AI applications are being used to deliver direct diagnoses of frequent conditions like the flu or distinguish less common conditions that can be mistaken for other conditions. With the use of a branch of machine learning called deep learning, visual recognition software is being used to advance diagnostics within subfields like radiology, pathology, dermatology, and ophthalmology. For example, earlier this year, the FDA approved the IDX-Dr system, the first autonomous diagnostic system that uses medical imaging to detect diabetic retinopathy, a condition when high blood sugar levels damage blood vessels in the retina. Furthermore, these highly accurate diagnoses are available when experts disagree over the interpretation of results, saving resources associated with additional testing.
AI is also being increasingly used in prognosis, or the study of the progression of disease based on an individual’s symptoms. Machine learning applications are able to differentiate and focus solely on information that is relevant for tracking particular chronic health conditions, like diabetes and muscular dystrophy. Given several streams of information such as blood pressure and auditory capacity, machine learning is able to identify variables that differ between certain health conditions. By distinguishing what is relevant, AI makes prognosis more efficient and reliable.
At The Institute of Cancer Research in London and the University of Edinburgh, a team of scientists are applying these machine learning capabilities towards cancer detection. Researchers have developed a machine learning technique called REVOLVER to make predictions about the evolution of future tumors based on repeating patterns of DNA mutations in existing tumors. With the knowledge gained from this AI tool, personalized interventions can occur at an earlier stage, overriding cancer’s main advantage — unpredictability.
In addition to diagnosis and prognosis techniques, artificial intelligence is also being incorporated in the day-to-day functions of medical practice. By analyzing rapidly growing amounts of patient history data, AI is able to extract clinically relevant information, facilitating the decision-making of health professionals. For example, with consensus algorithms developed by experts and health history data, computers can review and establish treatment alternatives like the most appropriate “cocktail” of chemotherapy drugs.
Furthermore, robotic tools controlled by AI have been used for a spectrum of surgical procedures, from tying knots and closing wounds in keyhole surgeries to removing tumors very close to the sensitive spinal region. According to Forbes, the increased consistency and lower error rate among robotic assisted surgeries has reduced hospital stay lengths by 21%. In fact, a study conducted by Mazor Robotics found that 379 orthopedic surgery patients with AI-assisted robotic procedures had five times fewer complications than regular surgery patients. Some heart surgeons use Heartlander, a mini robot that enters chest incisions to perform heart mapping and surface-level therapy.
AI applications help streamline healthcare administrative tasks. Technologies like voice-to-text transcriptions can help order tests, prescribe medications and write chart notes. An example is Nuance, a company that claims to slash the documentation time of health records by 45% and improve reporting quality by 36% through AI-powered solutions. Computer-assisted physician documentation solutions like Nuance offer guidance during the documentation process through consistent recommendations that facilitate natural workflow. In addition to helping health professionals save time on routine tasks, the automation of administrative tasks ensures that patients receive an accurate clinical history and proper guidance.
Artificial intelligence is also applied within telemonitoring, or the remote monitoring of patients that are not at the same location as the health care provider. The Berkeley Telemonitoring Project, one of UC Berkeley’s larger AI research initiatives, has developed a platform for devices that can collect data from a patient, deliver it to a remote server, and automatically analyze it. The analyzed data is then viewable by a healthcare provider, who can promptly offer medical feedback and interventions to the patient.
According to Dr. Daniel Aranki, a researcher in the UC Berkeley Electrical Engineering and Computer Sciences department and the executive director of the Berkeley Telemonitoring Project, the project began by addressing the rate of hospital readmission among Medicare beneficiaries with congestive heart failure. In collaboration with Northwestern Hospital in Chicago, 50 patients with heart failure were monitored for three months and assessed for certain risk factors.
“Given the cost, trauma caused for patients, and the sheer amount of people affected by heart failure, any improvement— even by a few percentage points — was considered significant,” Aranki explained. “Telemonitoring warrants a unified framework and that’s why we started the Telemonitoring Project in 2012 to create something more systematic. We started by looking at heart failure and then we ran more studies on other applications like the fitness of marathon runners.”
Applications of artificial intelligence in telemonitoring extend beyond projects pioneered by UC Berkeley. Virtual nursing assistants are now becoming available 24/7 to provide quick answers to basic questions and monitor patients, potentially saving the healthcare industry $20 billion annually. For example, San Francisco-based virtual nurse assistant Sensely recently garnered $8 million in funding to expand and keep patients and health care providers in better communication in between office visits, thereby reducing hospital readmission.
Despite the promise of AI to redefine the nature of healthcare, it’s important to recognize that efficiency and lowered operating costs are paired with some difficult realities. The use of AI in healthcare raises ethical questions about who is liable for machine error, inherent biases in the development of AI systems, and the privacy of sensitive medical data. Furthermore, a loss of jobs and human empathy will inevitably accompany additional AI integration in the healthcare sector.
It is difficult to foresee whether these concerns will give rise to yet another AI winter, but it is undeniable that AI technologies will disrupt the healthcare system as we know it. According to Aranki, “The machine really has no wisdom. It knows which strategies work better, but it does not know why. You can still learn from it, but you cannot ignore established knowledge gained from thousands of years of studying the human body.”
A question remains that only the future can answer — will the wave of artificial intelligence in healthcare stand the test of time and move forward, or will it simply die down?