Why AI Adoption Is Critical to Fix the US Healthcare Crisis

Michael Howell, Google's Chief Health Officer, welcomed Bob Wachter, Professor of Medicine and Chair of the Department of Medicine at the University of California, San Francisco. Wachter is a highly influential figure in healthcare, an elected member of the National Academy of Medicine, and is credited with coining the term "hospitalist." He is the author of five books, including "The Digital Doctor" and his latest, "A Giant Leap: How AI is Transforming Health Care and What That Means for Our Future."

The Dire State of US Healthcare

Wachter began by quoting from his new book, highlighting the paradoxical nature of US healthcare: it delivers miracles in cutting-edge and intensive care (transplants, cancer immunotherapy, gene editing) and is staffed by caring, dedicated, and well-trained individuals. However, the system itself is described as "a headache wrapped in red tape inside the nightmare that Franz Kafka himself might have dreamed up while on hold with the insurance company." This stark contrast underscores the desperate need for transformation.

He emphasized that the healthcare system's shortcomings are not due to a lack of effort or skill from its professionals. Unlike the pre-ChatGPT era where Google Search was considered excellent, no one in healthcare believes the current system is perfect. It is deficient in almost every aspect: quality, safety, convenience, equity, access, and cost. Wachter invoked Biden's quote, "Don't compare me to the Almighty, compare me to the alternative," to illustrate that the alternative—the current healthcare system—is profoundly inadequate. Therefore, an acceptance of the status quo, particularly regarding the slow adoption of AI, is not a neutral position but a negative one.

Shocking Statistics Underpinning the Crisis

Wachter presented several statistics to further illustrate the severity of the problem:

• Quality of Care: Evidence-based therapy, where the correct course of action is known, is delivered only about half the time. Patients receive it even less frequently due to factors like misunderstanding, affordability, or insurance issues. The average time from discovering the right treatment to implementing it is 17 years.
• Patient Safety: Medical mistakes cause between 50,000 to 200,000 deaths annually. This is often compared to a large airplane crashing every day, an event that would immediately halt air travel until fixed.
• Access: Finding a primary care doctor or mental health professional is incredibly difficult, even for those with resources and good insurance. Satisfaction with primary care is largely limited to concierge practices, where patients pay significant extra fees.
• Equity: Healthcare quality and longevity are significantly worse for poor and minority groups, with life expectancies decades shorter compared to wealthy white individuals.
• Cost: US healthcare costs are projected to reach $6 trillion this year, approximately 20% of the GDP. This exorbitant cost does not translate to a high-quality product, and the financial burden is increasingly shifted to individuals through out-of-pocket payments or lack of insurance.
• Administrative Burden: About one-third of healthcare spending goes towards administrative tasks, paperwork, and outdated technologies like fax machines. This administrative overhead adds little to no value for patients.

Despite these systemic failures, the US excels in high-end, complex procedures like lung transplants. However, the basic, everyday care that profoundly impacts longevity and quality of life is where the system falters.

The Evolution of Technology in Healthcare: From Paper to EHRs

Wachter discussed the history of computers in healthcare, noting that digitization lagged significantly behind other industries like finance, travel, and retail. In 2008, fewer than one in ten American hospitals had an electronic health record (EHR). Before EHRs, patient information was scribbled on paper, prescriptions were often indecipherable, and records were physical binders that had to be manually transported or faxed.

The widespread adoption of EHRs around 2008-2009 was primarily driven by federal incentives, including $30 billion in stimulus money during the Great Recession. Wachter's previous book, "The Digital Doctor," chronicled this transition from paper to digital, a process he initially approached with great optimism.

The Disappointment of EHRs and the Dawn of AI Optimism

Wachter's optimism for EHRs stemmed from his prior work on medical mistakes, many of which were caused by lost or illegible paper records. He envisioned EHRs eliminating handwriting errors, making patient charts universally accessible, and embedding evidence-based guidelines to reduce the 17-year lag in implementing best practices.

However, the reality was different. "The Digital Doctor" was a "grumpy" book because EHRs largely failed to deliver on their promise. Clinicians universally disliked them, finding them cumbersome and poorly designed. Wachter realized that while EHRs were a necessary foundation for digitizing data, they were not the ultimate solution. A significant portion of medical information, such as patient histories and medical literature, remained unstructured and not easily computable.

The advent of large language models (LLMs) and generative AI changed his perspective. Wachter, initially skeptical about writing another book on technology, was convinced by his wife and publisher. His research for "A Giant Leap," involving interviews with 110 people, revealed how AI could solve many of the problems that EHRs couldn't. AI offered the potential to:

• Make computerized decision support truly effective.
• Reduce the burden of documentation, allowing clinicians to focus on patients.
• Reintroduce humanism into medicine by freeing doctors from constant typing.

This led to a newfound optimism, a feeling he described as more enjoyable than the "grumpy, hopeless" sentiment associated with the EHR experience.

AI's Impact on Medical Education

Wachter explored the implications of AI for medical education, drawing insights from NYU, a leader in integrating AI with medical training. Key challenges and considerations include:

• Smarter or Dumber? A central question is whether access to powerful AI tools makes trainees smarter or dumber. Emerging data suggests it can make them dumber if not used thoughtfully.
• Governance: Academic medical centers face a tension between deploying AI tools for patient care and business efficiency versus ensuring trainees develop fundamental skills before relying on AI. For example, should trainees learn to write notes or review charts manually before using AI scribes or summarization tools?
• Human in the Loop: AI tools are not 100% accurate. If they were right only 50% of the time, they'd be useless. If 100% accurate, human intervention might be detrimental. The current reality is that AI is "right often enough to be useful and wrong often enough that you do want a human to be the final arbiter," especially in high-stakes situations like the ICU. Training must focus on understanding AI's limitations and how to effectively be the "human in the loop."
• Vigilance: Humans are poor at maintaining eternal vigilance when a technology they trust is mostly correct. This "deskilling" of critical thinking is a significant concern.
• Essence of Diagnosis: There's debate on how much core diagnostic knowledge trainees need if AI can provide answers. Wachter argues that understanding enough to ask the right questions, identify salient data points from a vast amount of information, and critically evaluate AI's output is crucial. Over-reliance on AI could lead to medical students remaining novices.
• Expert vs. Novice Use: Experts use AI tools differently than novices. Experts know what information to input and how to interpret the results. Novices, lacking this foundational knowledge, can easily misuse AI. For instance, a patient using a chatbot might describe a "bad headache" instead of the critical "worst headache of your life," leading to dangerously incorrect advice.

While some rote memorization (like the Krebs cycle) might be removed from the curriculum, Wachter believes it would be a mistake to significantly reduce the cognitive work required in medical training.

Areas of Improvement and Caution with AI in Healthcare

Meaningful Improvements:

• AI Scribes: Wachter uses an AI scribe in his clinical practice, which, despite not being perfect, allows him to maintain eye contact with patients and improve the human connection.
• Chart Summarization: AI tools can summarize lengthy patient records (e.g., 700 pages), making it feasible for clinicians to quickly grasp essential information.
• Administrative Tasks: AI is being used to streamline billing and other administrative processes, potentially reducing the one-third of healthcare costs currently spent on these tasks.
• Clinical Decision Support (Open Evidence): Tools like "Open Evidence" (similar to GPT or Gemini but trained on medical literature) are proving highly valuable. They allow clinicians to input complex patient cases and receive evidence-based answers with references, acting as a "curbside consult" that was previously unavailable. This is a significant leap beyond previous tools like UpToDate, which could only answer simple, structured questions.

Watch-Out Areas:

• Human-AI Dyad: The interaction between humans and AI is complex. While AI may not be the final arbiter, humans can become complacent or make errors when reviewing AI-generated suggestions. The AI's confidence level is also not yet transparent.
• Undue Trust: AI's ability to "act like a who" (human) rather than a "what" (technology) can lead to an "undue amount of trust," making critical evaluation difficult.
• Job Replacement: While Wachter believes AI won't immediately replace many clinician jobs due to unmet needs, the anxiety around job security is real, especially in unionized healthcare environments. The example of radiology, where initial fears of job loss proved unfounded, suggests that AI will augment rather than fully replace human roles.
• Cognitive Deskilling: As AI becomes embedded in processes and suggests diagnoses, tests, and treatments, there's a risk that clinicians will "turn their brain off." This could obscure the complex judgment, ethics, quality, and cost trade-offs inherent in medical decisions.
• Patient Use of AI: Patients are increasingly using AI tools, which can democratize access to information but also create conflict. AI tools for patients need to be designed differently, prompting for critical information (like the "worst headache of your life" example) before offering advice. The doctor-patient relationship will need to be renegotiated as patients arrive with AI-generated information, some of which may be accurate and some dangerously wrong.

The Risk of Going Too Slow

Wachter concluded by referencing a conversation with Gene O'Connell, CEO of the Mayo Clinic, who stated that the risk of going too fast with AI is less than the risk of going too slow. This sentiment echoes Wachter's initial point: the status quo in healthcare is unacceptable, making rapid, thoughtful adoption of AI a necessity rather than a luxury.