As the complexity and cost of health care increase, many health care institutions have turned to information technology as a means of improving patient care. Electronic medical records, patient portals, digitized medical devices, and even wearables are becoming more broadly used. These systems have been largely transactional, but artificial intelligence (AI) systems that are capable of machine learning go beyond traditional medical transactions and record-keeping to analyze data, make decisions, and exercise judgment.
Because AI tools perform tasks previously performed by humans, they raise concerns about large-scale job loss in health care and other industries. Our view, however, is that “augmentation” of human labor is more likely than large-scale automation.
The Potential of Cognitive Technologies in Medicine
Cognitive technologies are being been introduced in health care in part to reduce human decision-making and the potential for human error in providing care. Medical errors are the third leading cause of death in the United States, but they are not generally due to inherently bad clinicians. Instead, they are often attributed to cognitive errors (such as failures in perception, failed heuristics, and biases), an absence or underuse of safety nets and other protocols, and unwarranted variation in physician practice patterns.
The use of AI technologies promises to reduce the cognitive workload for physicians, thus improving care, diagnostic accuracy, clinical and operational efficiency, and the overall patient experience. While there are understandable concerns and discussion about AI taking over human jobs, there is limited evidence to date that AI will replace humans in health care. For example, numerous studies have suggested that computer-aided readings of radiological images are just as accurate (or more so) than readings performed by human radiologists.
But such systems are not yet in broad use, and, when they are used, they serve as a “second set of eyes.” We know of no radiologists who have lost their jobs from this form of automation. AI technologies such as IBM Watson have excited observers with their potential to treat cancer, but they don’t seem to have replaced any oncologists and, for that matter, there have been no rigorous examinations of their impact on patients. Sedasys, a semi-automated system for administering the anesthesia drug Propofol, met with poor sales and resistance from anesthesiologists and was withdrawn from the market. AI technologies may automate some medical tasks in the future, but few if any jobs have been fully computerized thus far.
Instead of large-scale job loss resulting from automation of human work, we propose that AI provides an opportunity for the more human-centric approach of augmentation. In contrast to automation, augmentation presumes that smart humans and smart machines can coexist and create better outcomes than either could alone. AI systems may perform some health care tasks with limited human intervention, thereby freeing clinicians to perform higher-level tasks.
Five Approaches to Augmentation
Augmentation blends the power of AI with the sophistication of human perception, empathy, and experience. In augmentation situations, the knowledge worker can either lead or support the automated decision tool. In robotic surgery, for example, the human surgeon has thus far played the lead role in surgical decisions and actions. In the future, however, enhanced diagnostic capabilities from technologies such as deep learning may mean that, for routine radiology and pathology tasks, machines may play the lead role.
We suggest there are five alternative roles for human clinicians in which different forms of augmentation take place:
Process Design Role:
Clinicians or managers can play a process design role in which they focus on how AI supports the process in question. This role involves (1) determining the overall role and value of a smart machine in a clinical process and (2) proposing and evaluating new AI technologies. In anesthesiology, for example, now that the Sedasys system has been withdrawn from the market, companies are focusing on technologies that will allow the anesthesiologist to focus on other technologies providing greater insight about their patients. Technologies such as Perioperative Point-of-Care Ultrasonography have been effective for postoperative bladder retention measurements by anesthesiologists, giving them a much broader view of the patient’s recovery after surgery.
Human Capability Role:
Clinicians can adopt a human capability role in which they primarily employ uniquely human skills such as empathy and interpersonal intelligences. In this role, they are choosing not to compete with the purely rational cognition and computation available from machines. For example, in radiology, the capabilities of deep learning for image analysis and diagnosis may eventually lead radiologists to focus on other types of tasks that AI can’t perform. As one study showed, empathy and emotional intelligences may be related to improved patient care. In that study, the inclusion of a patient’s photograph with the medical imaging results led to more meticulous image readings by the radiologist and revealed medical information regarding physical signs of disease and suffering.
Human clinicians may play the role of colleague alongside smart machines by evaluating the machines’ immediate outputs, determining if the data seem reasonable, and using this information to augment or inform their own judgments. For example, in oncology, genetic sequencing and other approaches to precision medicine have greatly increased the complexity of cancer diagnosis and treatment options. For breast cancer, approximately 75 drugs have been approved to prevent or treat the disease, alone or in combination, and treatment options can be informed by a patient’s genomic and metabolic function data.
Given this high level of informational complexity, human oncologists need insights and recommendations from intelligent machines. Tools such as IBM Watson are currently being trained to supply this information at cancer centers such as Memorial Sloan Kettering. The goal for these systems is not to replace physicians’ diagnostic capabilities, but rather to augment them.
Clinicians can fill a niche role for which no technology has yet been developed and likely will not be developed because it would not be feasible or economical to do so. In pathology, for example, automated systems can already perform standard screenings such as Pap smears. More esoteric tests, however, are performed too rarely to be automated soon. One such niche is pathology in orphan disease testing and drug development. Because there are >7,000 distinct orphan diseases, each of which afflicts <200,000 individuals, it is unlikely that technologies will usurp pathologists in this niche.
Clinicians can also play a development role with respect to AI technologies that other clinicians will use. In this role, they may work as researchers or in collaboration with AI vendors. In robotic surgery, for example, surgeons have been actively involved in the development of the next generation of surgical robots that may operate without human intervention. Recently, researchers at the University of California, Berkeley, with surgeons as their advisors, have developed robots that can perform entire operations — at least the repetitive and less-complex ones. Several surgeons at the Sheikh Zayed Institute for Pediatric Surgical Innovation have also worked on a system that does better than human surgeons at stitching soft tissues but still requires some human supervision.
Overcoming Barriers to Automated Decision-Making Through Augmentation
When health care organizations adopt a philosophy of augmentation, they can calm legitimate concerns about job insecurity and errors of automated decision-making. They also can encourage clinicians to learn new skills such as informatics and to explore new AI technologies to make augmentation roles more feasible and comfortable.
Technology diffusion in health care is often relatively slow, but we believe that augmentation approaches may speed up the rate of diffusion for AI. If doctors are working closely alongside the technology, both hospitals and payers may find that the adoption of AI technology will align with their desired cultures of measurement and accountability. Augmentation also can involve careful specification of the limited circumstances in which machine-led tasks are appropriate, which may ease approvals by organizations concerned with safety. The American Society of Anesthesiologists, for example, dropped its opposition to the Sedasys system after the manufacturer (Johnson & Johnson) limited its use to colonoscopies.
In some cases, AI tools can be combined with existing technologies to further augment clinicians in preventing medical errors. Start-ups like MedAware are integrating AI capabilities with electronic health-record systems to prevent prescription errors. Automated drug dispensing cabinets are already in fairly wide use within hospitals, decreasing the likelihood of drug errors. AI-based diagnosis or treatment recommendations also can reduce other types of medical errors, although thus far such recommendations are subject to clinician approval and adoption.
Addressing the Interests of the Health Care System, Physician, and Patient
By adopting an augmentation approach to implementing AI-enhanced care decisions, health care organizations can more effectively address the interests of multiple stakeholders. Clearly, the health care delivery system overall can benefit from the greater efficiency and effectiveness that smart machines can provide. In addition, physicians can be freed up to deal with nonstandard diagnosis and treatment issues and to advise patients on improving their health behaviors. Finally, and perhaps most importantly, both technologically savvy patients and patients in need of empathy can benefit from augmentation by seeking out health care organizations where the knowledge, speed, and low cost of AI are effectively combined with the caring and emotional concern of human beings.