Machine healing

During his second year as a medical student in the 2000s, Adam Rodman utilized the library for a patient whose health issue had baffled physicians. Rodman explored the catalog, duplicated research papers, and provided them to the care team.

“It made a significant impact on that patient’s treatment,” Rodman remarked. “Everyone said, ‘This is wonderful. This is evidence-based medicine.’ But it required two hours. Today, I can achieve that in 15 seconds.”

Rodman, currently an assistant professor at Harvard Medical School and a physician at Beth Israel Deaconess Medical Center, now carries a medical library in his pocket — a smartphone application developed following the launch of the large language model ChatGPT in 2022. OpenEvidence — partially created by Medical School faculty — allows him to inquire about specific illnesses and symptoms. It searches through medical literature, generates a summary of discoveries, and provides the most relevant sources for further examination, supplying answers while Rodman is still in front of his patient.

“We say, ‘Wow, the technology is indeed formidable.’ But what steps do we take to genuinely enact change?”

Adam Rodman

AI in various forms has been employed in the field of medicine for decades — but not in this manner. Analysts forecast that the integration of large language models will transform healthcare. Some liken the potential influence to that of decoding the human genome or the advent of the internet. The effects are anticipated to be evident in doctor-patient interactions, the workload of physicians, the administration of hospitals and practices, medical research, and education in medicine.

Most of these outcomes are expected to be beneficial, enhancing efficiency, diminishing errors, alleviating the nationwide shortage in primary care, contributing data more thoroughly to decision-making, lessening administrative loads, and creating opportunities for longer, more substantial one-on-one interactions.

Nonetheless, there are substantial worries as well.

Current datasets frequently embody societal biases that perpetuate disparities in access and quality of care for underprivileged groups. If unaddressed, these data risk entrenching existing biases into increasingly powerful AI technologies that may

increasingly shapes the manner in which healthcare functions.

Another significant concern, specialists argue, is that AIs are still susceptible to “hallucination,” fabricating “facts” and presenting them as if they are genuine.

Moreover, there’s a risk that the medical field may not be sufficiently daring. The newest AI has the capability to overhaul healthcare from the ground up, but only if granted the opportunity. Misguided priorities — excessive respect for established interests, a concentration on profit rather than wellbeing — could easily diminish the AI “revolution” to a lackluster attempt at superficial modifications.

“I believe we find ourselves in a peculiar position,” Rodman commented. “We proclaim, ‘Wow, the technology is incredibly potent.’ But what actions do we take to truly transform situations? My concern, as both a practitioner and a researcher, is that if we don’t envision grand changes, if we don’t re-examine how we have structured medicine, things may not evolve significantly.”

Reinforcing the ‘unstable structure’

Five years ago, when inquired about AI in healthcare, Isaac Kohane expressed exasperation. Teenagers engrossed in social media applications were better prepared than many medical professionals. The current scenario is starkly different, he claims.

Kohane, head of the Medical School’s Department of Biomedical Informatics and editor-in-chief of the New England Journal of Medicine’s new AI initiative, describes the capabilities of the latest models as “astonishing.” To underline this point, he recalled having an early preview of OpenAI’s GPT-4. He evaluated it using a complicated case — a child born with indistinct genitalia — that could have perplexed even a seasoned endocrinologist. Kohane consulted GPT-4 about genetic origins, biochemical processes, subsequent steps in the evaluation, and even the appropriate information to share with the child’s parents. It performed excellently.

“This advanced language model was not specifically designed to function as a physician; it’s merely trained to forecast the following word,” Kohane remarked. “It could articulate as fluently about wine pairings with a vegetarian feast as it could diagnose a complicated patient. It was genuinely a remarkable leap from anything that anyone in the realm of computer science, who was honest with themselves, would have anticipated in the upcoming decade.”

And just in time. The U.S. healthcare system, repeatedly criticized as expensive, inefficient, and excessively focused on treatment over prevention, has been revealing weaknesses. Kohane, reflecting on a faculty member new to the department who couldn’t locate a primary care doctor, is weary of observing these issues firsthand.

“The healthcare system, which I have long claimed is flawed, is broken in quite apparent ways in Boston,” he stated. “People fret about equity issues with AI. I’m here to assert we are already facing a significant equity challenge today. Unless you have strong connections and are prepared to spend literally thousands more for concierge services, locating a timely primary care appointment will prove challenging.”

Initial fears that AI would supplant doctors have shifted to the understanding that the system requires both AI and its human workforce, Kohane indicated. Collaborating nurse practitioners and physician assistants with AI is one of several encouraging possibilities.

“It is no longer a discussion about, ‘Will AI displace doctors,’ but rather, ‘Will AI, together with a group of clinicians who may not resemble the physicians we are accustomed to, help stabilize the unstable structure that is organized medicine?’”

Creating the ideal assistant

The manner in which LLMs were introduced — to everyone simultaneously — hastened their assimilation, Kohane remarks. Physicians promptly began experimenting with monotonous yet vital tasks, such as composing prior authorization requests to insurance providers justifying the necessity of certain, typically costly, treatments.

“People simply embraced it,” Kohane noted. “Doctors were actively communicating on social media about the time they were conserving.”

Patients engaged as well, seeking virtual second opinions, similar to the child whose persistent pain was erroneously diagnosed by 17 different doctors.

over a span of three years. In the extensively covered incident, the mother of the boy inputted his medical records into ChatGPT, which proposed a condition previously unmentioned by any healthcare professional: tethered cord syndrome, wherein the spinal cord adheres within the vertebral column. When the individual moves, instead of gliding smoothly, the spinal cord is stretched, leading to discomfort. A neurosurgeon later verified the diagnosis and rectified the anatomical irregularity.

One of the anticipated advantages of utilizing AI in clinical settings, naturally, is to enhance the proficiency of doctors during the initial assessment. Increased and quicker access to patient histories, suggested diagnoses, and additional information is projected to elevate physician capabilities. Nonetheless, considerable challenges persist, as a recent study indicates.

A study published in JAMA Network Open in October contrasted diagnoses made by a single physician, a physician utilizing an LLM diagnostic instrument, and the LLM independently. The findings were unexpected, revealing negligible enhancement in accuracy for the doctors employing the LLM — 76 percent compared to 74 percent for the solo physician. More astonishingly, the LLM alone performed best, achieving a score 16 percentage points greater than the physicians working independently.

Rodman, one of the senior authors of the paper, noted that while it might be tempting to conclude that LLMs offer limited assistance to physicians, it is crucial to analyze the results more thoroughly. He stated that only 10 percent of the doctors were experienced users of LLMs prior to the study, which occurred in 2023, with the remainder receiving merely rudimentary instruction. As a result, when Rodman later examined the transcripts, most utilized the LLMs for fundamental fact-checking.

“The optimal way a physician could currently use it is for a secondary opinion, to challenge their own conclusions when faced with a complex case,” he remarked. “How might I be mistaken? What am I overlooking? What supplementary questions should I engage with? These are the strategies, we understand from psychological research, that complement human thought processes.”

Among the additional prospective advantages of AI is its ability to enhance the safety of medical practices, as stated by David Bates, co-director of the Center for Artificial Intelligence and Bioinformatics Learning Systems at Mass General Brigham. A recent investigation conducted by Bates and his team revealed that as many as one in four consultations at Massachusetts hospitals result in some form of patient injury. Numerous such occurrences can be traced back to negative drug-related incidents.

“AI should be capable of identifying medication-related problems and recognizing them much more precisely than our current capabilities allow,” commented Bates, who additionally serves as a professor of medicine at the Medical School and of health policy and management at the Harvard T.H. Chan School of Public Health.

Another possibility arises from AI’s increasing proficiency in a routine area: note-taking and summarization, as noted by Bernard Chang, dean for medical education at the Medical School.

Systems for “ambient documentation” are expected to soon be capable of listening in on patient consultations, documenting everything that transpires, and generating an organized clinical note in real-time. As symptoms are discussed, the AI can propose diagnoses and treatment strategies. Subsequently, the physician can assess the summary for its accuracy.

Automating notes and summaries would serve healthcare professionals in multiple ways, Chang indicated. It would alleviate the data entry burden that is often cited as a contributor to burnout, and it would transform the doctor-patient dynamic. One of the most common grievances from patients regarding office consultations is seeing the physician focused on the computer, asking questions and noting the responses. By being relieved from the note-taking chore, physicians could interact face-to-face with patients, paving the way for deeper connections.

“It may not be the most extraordinary application of AI,” Chang remarked. “We’ve all witnessed AI accomplish something and thought, ‘Wow, that’s impressive.’ This, however, is not one of those instances. Yet, this program is being tested in various outpatient practices nationwide, and the initial outcomes are very encouraging. Physicians who feel overwhelmed and exhausted are beginning to express, ‘You know what, this tool is going to assist me.’”

The bias dilemma

Despite their capabilities, LLMs are not yet suitable to operate independently.

“The technology does not possess the adequacy to reach a safety threshold wherein a knowledgeable individual is not required,” Rodman remarked. “I comprehend where issues may have arisen. I can take an additional step with the diagnosis. I am capable of doing this because I learned from challenging experiences. During my residency, numerous mistakes were made, yet I gained knowledge from them. Our present system is remarkably inefficient, but it cultivates critical thinking. When medical students engage with systems that can automate those processes — even if they tend to outperform humans on average — how will they acquire knowledge?”

Healthcare professionals and researchers also express concerns regarding inaccurate information. Predominant data bias originates from biomedicine’s foundations in affluent Western countries, whose scientific narrative was predominantly shaped by white males studying white males, states Leo Celi, an associate professor of medicine and a physician in the Division of Pulmonary, Critical Care and Sleep Medicine at Beth Israel Deaconess Medical Center.