Practical Musings About the State of Artificial Intelligence and Cognitive Computing
Marzieh Nabi, Ph.D. is the director of innovation commercialization at PARC, a Xerox company, with a technical background that encompasses AI and different aspects of systems engineering (control, optimization, networked dynamics systems, robotics, and flight dynamics). She is interested in applications of these tools in societal problems such as energy, transportation, aerospace, multi-agent and autonomous systems, and healthcare.
Artificial Intelligence in Healthcare: How valid is the hype?
We hear a lot about artificial intelligence (AI), machine learning (ML) and the potential of these technologies to turn medicine and the healthcare business on its head. Embedding decision support systems into the everyday work of clinicians, for example, can help improve diagnoses, suggest efficient treatment plans and enhance quality of life.
But deciphering AI hype from reality is challenging. The full implementation of AI and cognitive computing techniques in healthcare, by some accounts, could take 50 years – at the very least one to two decades. Why? The reasons include both technical and business challenges.
The Top 10 Technical Challenges
Today, healthcare organizations possess the data. Typically, they have limited resources to analyze it and little incentive to share it. The quality of AI algorithms, however, depends on the caliber of the data. The more high-quality data AI scientists have to work with, the better the algorithms behave.
Patient privacy and security is, understandably, a critical matter to hospitals and clinics. Implementing anonymization algorithms and other security protocols, along with patient consent, would ease some of these constraints. Improving communication with patients and family could also assist in data being more widely shared. It’s hard to imagine people would think twice when told data could be used to improve the overall quality of care.
Although these issues are challenging, we are starting to see how competition and the increasing need to provide better care is forcing to open the gates, allowing data scientists better access to the vast collections of data currently collected by these organizations. Open genomic data repository at MD Anderson Cancer Center is one example.
A second challenge related to the data is the fidelity, which refers to both variation and age of data. Medical data is heterogeneous, capturing many layers of complexity. It can include clinician notes, biological pathways, genomics, imaging, lab tests, diagnoses, procedures and prescriptions. Harvesting insights requires the design of data schemes that collect different data sets into a common database. This is a complex, technical challenge that still requires years of work.
Linking the AI and medical communities is another barrier. Any advancement in applications of cognitive computing in healthcare requires close collaboration. Open dialogue helps the AI community learn the medical community’s goals and agendas, workflow and how to best design the type of decision support that will be most accepted and used by clinicians. Two-way communication also helps the medical community gain knowledge about the potential and limitations of AI.
But the challenges go beyond data sharing and collaboration. Designing a decision-support system that leverages patient observational data to assist medical experts in their decision-making processes requires a complex “man plus machine” collaboration. There are challenges in “man-machine” integration in different industries, and healthcare won’t be any exception.
The aim is to design systems that enhance human performance because success in healthcare hinges on human intellect – especially with complex health situations such as comorbidities (co-occurrence of multiple chronic conditions), different cancer types, different epidemiological outbreaks, etc.
To accomplish this, the first step is to create a model of the clinician’s performance and mental workload. Not much work has been done to characterize different medical practices, but medical data can help. The task allocation between a clinician, or a team of clinicians, and the machine is another significant aspect to be considered. Other challenges include interface, communication and visualization. Evaluation and experimental validation also are necessary to establish trust between human experts and the new machines.
We must think critically about where new processes and new decision-support systems will fit into the workflow. Success will be determined by how well a system manages the workflow to accomplish better quality and outcome without adding more tasks to health providers’ “to do” list. AI and ML-driven technologies can help providers apply similar lean engineering principles used to remove waste and maximize efficiency in the automobile and other industries.
Medical workflow is distinct in different-size medical practices. Changing the workflow to improve quality and outcome is a crucial concern in healthcare. While clinicians and nurses are overwhelmed with different processes and guidelines, new applications and processes are trying to catch their attention and embed them within the workload.
The Top 5 Business Challenges
A well-known computing adage known as Amara’s law says we tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run. We are seeing this with ML and AI. According to research firm Gartner’s Hype Cycle for Emerging Technologies, for example, ML is at the peak of its “inflated expectation.” This means it is an important time for strategists and people in charge of creating business opportunities from AI/ML advancements in healthcare to understand the technology limitations. We need realistic expectations and roadmaps to move forward.
Step one is to acknowledge that healthcare is a fully regulated industry. This creates opportunities and imposes limitations. For example, while Medicare and Medicaid play a key business role, complying with regulations associated with these government programs requires both financial and time investment. This is where AI/ML can make a practical difference.
Step two is to sort out short-term and long-term goals. When it comes to use of technology, short-term plans involve thinking about low-hanging fruit. Solutions are less complex, and the business starts to see the return on investment sooner than later. But short-term agendas don’t often create sustainable business opportunities. Longer-term agendas delve into higher-risk and higher-reward problems. They require research, which takes time and money. It’s possible to design a system where the financial returns on short-term opportunities fund longer-term goals.
In the short term, for example, a variety of digital therapeutic technologies already are enhancing medicine. New sensors record whether and when a patient takes medication, and Apple’s CareKit software program allows patients to actively manage their own medical conditions.
Contrast this with the long-term challenge of elderly healthcare management. Creating technologies that can help meet both the physical and mental needs of these patients will require better artificial intelligence and robotic technology.
The healthcare ecosystem is complex with a diverse set of stakeholders, from patients to physicians, to nurses, to hospitals, to different-size clinics, to pharma, to operations, to insurance companies, to government. The current payment system in the United States adds another layer of complexity. All of this has challenges but also comes with great opportunity – if we can set realistic expectations and link the right players.
Despite technical and business challenges, many institutes, from government to academia to large and small corporations, have started to take steps toward realization of AI and cognitive computing in healthcare. We are working towards using deep learning to provide more insights on imaging and other patients’ and physicians’ data. We are not shy of embracing advanced tools such as Google Glass in healthcare workflow. Reading all medical literature in a short amount of time is now possible. This is an interesting era for AI and cognitive computing in healthcare. Even though we are far away from Dr.ai, I believe assist.ai is in our reach.
Our work is centered around a series of Focus Areas that we believe are the future of science and technology.
We’re continually developing new technologies, many of which are available for Commercialization.
PARC scientists and staffers are active members and contributors to the science and technology communities.