AI in Healthcare - Embrace the Wolf, Beware the Husky

Humans have always tended one another, human discovery has led the way, and we have developed tools to provide that care.? Our methods and instruments have changed and mostly improved by way of experience, practice, and ingenuity.? The one constant in the evolution of healthcare is humanity.? Humans will always need, suffer, and heal, and we will tend each other with the powers of discovery, learning, and improvement.

Our power to heal is evident every day in the pediatric ICU, and the legacy of this human ability is evident in excavated femur bones which healed in Tell Fara of the West Bank some 2,000 years ago.

I believe healthcare technology is 80 percent people, 15 percent process, and only 5 percent technology.? Technology is important, yet our people and the ways we work are far more important.? At CommonSpirit Health, our technology infrastructure empowers more than 150,000 healthcare professionals and other personnel to care for 20 million patients at more than 2,200 care sites, in 142 hospitals, in 24 states, every year.? Technology makes all of this possible, not on its own, but by serving those who care for other people.

While the introduction of AI generally and ChatGPT recently may have seemed sudden, the underlying technology has been available and in development for decades. Healthcare as an industry has not been ready to make full use of it until now.

One early and indispensable tool of health was the discovery of penicillin.? Nature took its course, but it required the power of human observation to understand what nature provided.? We had a powerful discovery.? We needed to convert our new knowledge into a treatment, produce that treatment, deliver the care, cure the patient, and teach others to carefully and reliably deliver the same result to patients in need.??

When Dr. Alexander Fleming discovered the antibacterial properties of penicillin in 1928, its broader use was not widely understood.? Its first human use was not until 1941, and, in that year, there was only enough penicillin in the world to briefly treat one patient.? His infection improved, but he died a few days later when the supply ran out.? By 1943, penicillin was being mass-produced for war use, and, by 1945, it was part of a standard of care for the general public.?

Similarly, AI was first developed as an academic concept in the 1950s and as a focus of military-funded research in the 1960s.? Within healthcare, we have been using rules-based expert systems and narrow AI applications for back-office automation for decades.? Only in the past ten years, however, has the technology advanced to the point at which we are applying it to clinical processes.? The path to healing and cure has not been without mistakes and setbacks.? Over time, both AI as a tool and our ability to employ it clinically will advance and improve alongside one another.??

In 2016, Karen Zack created the chihuahua versus blueberry muffin meme which has been seen by many as a fun challenge for AI visual recognition and classification.

That same year, researchers at the University of Washington published the paper “Why Should I Trust You?”: Explaining the Predictions of Any Classifier (Ribeiro et al.) in which they challenged the believability of black box machine learning models.? The most compelling example from the paper came from the wolf versus husky challenge which revealed that the AI model was predicting the breed of the canine not based on the characteristics of the animal, but on whether snow was visible in the picture.?

We are technology generations beyond these examples, but they highlight a fundamental issue of AI adoption in healthcare.? Our industry can embrace advances that robotic process automation (RPA), machine learning (ML), and large language model (LLM) AI can bring to efficiency, staffing, cost savings, quality improvement, clinician unburdening, and patient satisfaction, yet AI cannot be given full clinical autonomy in the near future.? Clinicians must have the final say.

We should save human discernment for questions best decided by humans, and use tools for the parts where we fall short.? Used precisely, AI tools have the power to support humans in health, healing, and delivering care as well as identifying times of need–sometimes even before symptoms appear.??

The fuzzy edge of AI and LLM applications tend to receive a lot of focus, but only seeking ways to apply new algorithms overlooks the ways we are already actively using AI in healthcare today.? And while AI is currently thought of as a futuristic technology, its use is rapidly becoming standard in back-office systems and in healthcare process applications.??

We successfully use AI at CommonSpirit, running approximately 60 AI-based systems which support clinical and operations processes.? These are full-scale systems, not pilot projects, and they cover areas as wide-ranging as imaging, diagnostics, and billing.?

The stack above outlines the expanding use of AI in healthcare.? Most US health systems have been using robotic process automation in back-office processes since the late 2010s.? Like other industries, RPA has created efficiencies and savings in routine timekeeping and invoicing tasks.? Healthcare-specific RPA applications like health information management and patient financial services mirror those of customer satisfaction in other industries.

One of the best generalized applications of routine AI in healthcare is inbound call automation.? Unlike businesses with a single front desk or toll-free number, health systems have hundreds of patient-facing phone numbers including hospital switchboards, physician practices, and pharmacies.? Patients calling any one of those numbers at a health system have the reasonable expectation that they can be transferred to the correct department, and yet the health system employees answering the phone generally have neither knowledge of other departments nor access to a comprehensive directory.? Calling and asking for the hyperbaric wound care center could lead to numerous transfers unless the patient gets the right person with the right understanding on the phone.? Although the idea of having AI answering the phone may not seem patient-friendly, being quickly connected to the correct department is, in fact, good patient service.

Benefiting Employees

While we often focus on how AI can enhance accuracy and free clinicians of rote work so that they may devote more time to patient care, AI also has the capacity to offer ancillary support to all of our staff.? For example, a member of our staff shared how Ida, our CommonSpirit AI assistant, provided service-desk assistance.? Ida provided easy step-by-step instructions to set up a company email on a new phone.? This use is important, because it not only empowers employees across our enterprise to resolve administrative questions quickly, conveniently, and independently, it allows our service-desk staff to devote their time to solving more complex issues.? Staff satisfaction with Ida has increased to 88 percent, and we expect to see even more improvement.? Ida handles an average of 6,000 issues each month, and creates greater than $200,000 in efficiency and labor savings annually.

Benefiting Patients

It is important to emphasize that the role of AI is not limited to supporting and empowering clinicians but also supporting and empowering patients.? Patients can be more informed and autonomous in seeking, choosing, and receiving their care than at any other time in our history.????

We are implementing AI in our patient interactions as well.? Voice-driven AI can help patients who call CommonSpirit to schedule appointments.? Beyond scheduling, our newer AI program is able to use patient history to reduce cancellations and no-shows with better contact and reminders.? In Nebraska, we are already using this AI model to identify barriers to treatment and ensure patients receive care to reduce the likelihood of more critical illness and hospitalization.

In our Arizona hospitals, we have partnered with a company to roll out an AI-based tool which communicates with patients who come to our Emergency Departments (ED) by way of smartphone updates on wait times, the expected next step in their care, the status of their labs and imaging, and even a simple explanation of clinical fundings, all of which can be linked to family members to keep them informed in real time.

Benefiting Clinicians

As AI has improved, we have begun adopting it for clinical use.? A tenant of AI use in patient care is the need to ensure that it is free of algorithmic bias.? Whether the fault of the tool itself or the data on which it is trained, algorithmic bias is the risk that AI-based clinical processes reflect social inequities, use outdated decision models, and ultimately exacerbate existing health disparities.? At CommonSpirit, we surveyed our existing tools to ensure we had no embedded tools with algorithmic bias, and we continually evaluate all new technology we introduce to our system for the same risk.? Even more importantly, the benefit of being a health system which cares for 20 million patients across America is that the data sets we use to train our AI tools are large and reflect the makeup of the patients we serve.? Our diversity lends to our strength.

Like many health systems, we have begun rolling out tools to reduce the documentation burden for physicians and nurses.? We have partnered with an independent company on an application to reduce the burden of documentation for physicians and nurses.? Ambient scribing, driven by an AI tool embedded on the clinician’s smartphone and with the patient’s permission, translates clinical patient conversations into structured notes, populates critical fields, and initiates actions in patient EHRs.? And, while most patients are accustomed to seeing clinicians make notes during visits, ambient scribing can allow clinicians to engage even more directly with patients without sacrificing the accuracy of EHR notes.

Similarly, automated inbox management can save clinicians’ time by filtering and routing appropriate messages before providers even check the inbox of the EMR.? Ensuring that patient and clinician messages are routed to the correct recipients and proposing draft messages to physicians has the promise of reducing communications volume and the time of response to patients.

For real-time clinical care, AI is now instrumental in identifying sepsis (infection) and stroke.? Since 2016, we have used AI technology which scours patient medical records to identify patterns of sepsis risk.? Constant monitoring of the most subtle changes in a patient provides an added set of eyes to ensure a redundant fail-safe even beyond our existing best-proven processes.? CommonSpirit has used the AI sepsis algorithm to monitor 17.9 million patients, reduce an average of 33 hours of ICU time per patient, and, most importantly, save more than 1,600 lives.

We have also used a tool called Stroke Aware since 2019 to apply precision AI to read CT angiography to look for blood clots in the brain.? Here, AI uses a series of algorithms to immediately alert physicians to areas of concern which may be even too small for the human eye to detect.? The power of AI to flag potential causes of a stroke immediately allows a doctor to develop a care plan without losing a moment of the time so critical in treating and preventing stroke.?

At the top of this stack is automated diagnosis and action.? While AI which is advanced enough to provide quality care for entire populations is a goal, we are years from a point where we can entrust the care of patients solely to technology.? AI can supplement and suggest care processes, but clinicians with their patients need to be the ones who make the final choices.? This idea aligns with the November 2023 American Medical Association principles for AI use which state, in part, “Clinical decisions influenced by AI must be made with specified human intervention points during the decision-making process.”

There is always a clinician between AI and the patient.

Humanity is the beginning and end of healthcare

Ribeiro, Marco & Singh, Sameer & Guestrin, Carlos. (2016). “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations 97-101. 10.18653/v1/N16-3020