Mike Magee MD
Our third session of “AI Meets Medicine” will ask the question, “Has the future arrived?” As we will see, the clear answer is “yes,” but not without future risks. In the case studies that follow, we will see that our progress has been rapid, and touched nearly every branch of Medicine. But challenges remain because, as David Brooks recently noted, AI has three split personalities – the brain of a scientist, the drive of a capitalist, and the cautious heart of a regulatory agency.”
AI also has attracted an onslaught of aggressive tech entrepreneurs in an epic face-off that has peaked in the past five years. These billionaire oligarchs include Elon Musk with his new XAI venture, Sam Altman whose OpenAI has attracted $10 billion in Microsoft funding on the backs of ChatGPT, Mark Zuckerberg with Meta and is highly enabled metaverse googles, and Gunder Pichai who is driving Googles AI leader app, Gemini.
The top U.S. sector consumers are Energy, Health Care, Aerospace, Construction, and Supply Chain – with healthcare being #2 overall. This is a reflection of its’ outsize financial position in our economy, its enormous complexity, and its’ science, technologic, and research bases.
In the recent 2023 AI report generated by the Boston Consulting Group (BCG), more than 60 use cases for generative AI in health care were fleshed out. As the report stated, “Success in the age of AI will largely depend on an organization’s ability to learn and change faster than it ever has before. . . Deploy, Reshape, Invent.” Action steps included massive upskilling, new roles and budget allocation, technology platform development, genAI enabled assistants, reshaping customer service and marketing, massive efficiency gains (30% to 50%), and oversight of accuracy , safety and ethics.
With all of the above, it should come as no surprise that “Generative AI is projected to grow faster in health care than any other industry, with a compound annual growth rate of 85% through 2027.” And yet, the experts remain filled with anxiety. Consider the discussion between Open AI’s Sam Altman and his 2023 Podcast host, MIT Tech Engineer, Lex Fridman:
Lex: “You sit back, both proud, like a parent, but almost like proud and scared that this thing will be much smarter than me. Like both pride and sadness, almost like a melancholy feeling, but ultimately joy. . .”
Sam: “. . . and you get more nervous the more you use it, not less?”
Lex: “…yes.”
But at the same time, the opportunity to do great good seems irresistible. AI has the capacity to “parse through vast amounts of data; glean critical insights; build predictive models; improve diagnosis and treatment of disease; optimize care delivery; and streamline tasks and workflows.” … and we haven’t even mentioned scientific discoveries.
Let’s examine a few illustrative cases.
Case 1. is the discovery of the Covid vaccine. It is a “recent past reality” that none will soon forget. The crisis began on December 1, 2019 when a local person in Wuhan, China appeared at the hospital with fulminant pneumonia. By December 30, 2019, the local populace was in a panic when more and more citizens became ill with little explanation. On January 5, 2020, famed virologist Shi Zengli at the Wuhan Institute of Virology revealed the full viral genetic code of the infecting agent. Later that month local health officials revealed that 571 individuals had now been infected and several had died of pulmonary complications.
Word of the local epidemic had now begun to seep out despite the Chinese officials attempts to keep the health challenge silent. One local official was quoted saying, “It erupted too fast, and then there were just too many people infected. Without ventilators, without specific drugs, even without enough manpower, how were we going to save people? If you’re unarmed on the battlefield, how can you kill the enemy?”
On February 15, 2020, just 45 days after receiving the viral genetic code, the pharmaceutical company, Moderna, announced that they had created a “clinical-grade, human safe manufacturing, batch (of mRNA) shipped to health clinics for testing.” This surprised scientists inside and outside the company. To create vaccines requires testing against varied samples of mRNA’s created in the laboratory. Normally, this is a laborious process yielding around 30 samples a month. But the company had created an AI energized process capable of creating over 1000 samples in one month, and then found the sample that worked.
Dave Johnson PhD was the head of their AI project and later said, “We always think about it in terms of this human-machine collaboration, because they’re good at different things. Humans are really good at creativity and flexibility and insight, whereas machines are really good at precision and giving the exact same result every single time and doing it at scale and speed” The rapidity was the result of AI driven hyper-accelerated mRNA generation. AI was then used again to predict how best to structure the vaccine to maximize a protein production response in the body…or as the company says, “more bang for the biological buck.”
In the meantime, the U.S. was struggling inside and outside Washington, D.C. President Trump sowed confusion at his daily Press Briefings, leaving Vice President Pence, and Dr’s Tony Fauci and Deborah Birx in confused silence. States, cities, and municipalities closed schools and congregate business sites, and mandated use of masks. And death rates continued to escalate from what was now recognized as a worldwide pandemic.
In the meantime, Moderna, and competitor Pfizer, accelerated human testing, and on December 18, 2020 received “emergency use authorization” from the FDA Vaccine Advisory Committee. This was not a moment too soon, most would say. The death toll in the U.S. had already reached over 800,000, and projections of monthly fatalities ahead had reached 62,000. Looking back a few years later, lives saved worldwide as a result of the AI assist were calculated at 15 to 20 million.
Our second case study, AI assisted Facial Recognition Technology (FRT), on the surface may seem totally unrelated to the Covid vaccine story above, but there is in fact a link. FRT science in America has a rich and troubled history. The research modern era began in 1964 under the direction of William Woodrow Bledsoe, a information scientist in Palo Alto, CA.
An American mathematician and computer scientist, he created now primitive algorithms that measured the distance between coordinates on the face, enriched by adjustments for light exposure, tilts of the head, and three dimensional adjustments. His believe in computer added photography that could allow contact free identity had triggered an unexpectedly intense commercial interest in potential applications primarily by law enforcement, security, and military clients.
A half century later, America has learned to live under rather constant surveillance. On average, there are 15.28 cameras in the public space per capita nationwide. Each citizen on average is photographed unknowingly 238 times per week.
But how did Covid collide with FRT? The virus itself, according to one theory, was original constructed, with support from American funds (through DARPA, our military funding agent), by US trained Chinese virologist Shi Zengli at the Wuhan Institute of Virology. At the same location, DARPA, in pursuit of masked terrorists around the globe, had funded an AI assisted FRT study (during the pandemic masking period) to see if facial recognition might be possible, even with masks on in Wuhan. It was.
That finding in 2020 was of interest to front line hospital managers, looking for a hands-off method of registering and tracking admitted in-patients in facilities overwhelmed at the time with Covid patients. That opened up new possibilities for research. With each citizen’s facial data already on record, imagine if we allowed genAI loose on it. Let’s see what it can already do in the diagnostic arena, they said.
Over the next months, new AI-FRT applications came rolling out. The UC San Diego School of Medicine focused on eye-gaze research. Using patented “eye-tracking biomarkers,” they successfully identified “autism” in 400 infants and toddlers with 86% accuracy. A similar technology rapidly used “facial descriptors” as an entry vehicle to confirm “syndrome gestalts” in successful AI face-screening analysis of rare genetic disorders.
If possible in infants, why not seniors? In 2021, “Frontiers in Psychology reported “The effectiveness of facial expression recognition in detecting emotional responses to sound interventions in older adults with dementia.” Surely, hands off diagnosis of sub-clinical Alzheimers can’t be far beyond. The study results above were made possible because AI eliminated 98% of the required coding time. In a similar vein, compliance and accuracy of patient reporting (especially in the conduct of funded clinical trials) was an obvious fertile area. AiCure, a data app, offered “Patient Connect”, an AI driven facial surveillance to monitor ingestion of medication as directed in study protocols. Privacy advocates wondered, could a similar approach double check that an elder patient’s contention of 1or 2 alcoholic drinks a week was accurate?
If all this is beginning to make you uncomfortable, it may be because, as a nation, we’ve been down this path before and it didn’t end well. That back story for FRT began with a Cambridge based genius statistician named Sir Francis Galton. As one historian recounted, “In 1878 Galton began layering multiple exposures from glass-mounted photographic negatives of portraits to make composite images comprised of multiple superimposed faces. These photographs, he later suggested, helped communicate who should and should not produce children. This photographic technique allowed him to give enduring, tangible, visual form to ‘generic mental images,’ or the mental impressions we form about groups of people—better known today as ‘stereotypes’.”
Galton, whose cousin was none other than Charles Darwin, was more than familiar with “survival of the fittest.” And he was quite committed to excluding those troubled by poor health, disease and criminally from the “survivor’s list.” Elite leaders and the universities that had educated them agreed. The programs they created advanced the cause of eugenics, which by the 1920’s began to leave its mark on the law.
In one famous case, Buck v. Bell, Carrie and Emma Buck, mother and daughter, were involuntarily confined to a North Carolina Mental institution, without evidence of mental disease or support for claims they were “idiots” or “imbeciles” (legal terms at the time.) Carrie’s problems developed in her teens, when the nephew of her foster parents (her mother was already institutionalized) raped her. A normal, healthy child was born, and shortly thereafter, Carrie was separated from her child and confined with her mother in the mental institution. She was then hand selected by local politicians as a test case to see if they had the power to sterilize select citizens without their permission. In Buck v. Bell, the most famous Justice of that era, Oliver Wendall Holmes, decided that North Carolina did have that power and famously declared to the joy of eugenists everywhere, “Three generations of imbeciles are enough.” Carrie’s Fallopian tubes were subsequently tied.
After Hitler adopted the U.S. Eugenics Laws as he rebuilt Germany in his image in the 1930’s, U.S. officials began reconsidering their efforts on behalf of human purity. But at Yale, between 1940 and 1970, male and female students were routinely required to be photographed naked as part of a eugenics study. And in many states, eugenics laws remained on books into the 1970’s. It was only in 1993 that President Clinton offered a formal apology to all those injures on behalf of the nation.
Our third “AI in Medicine” Study focuses on AI-assisted diagnostic acumen. Just this week, the New York Times reported that Chat-GPT4 had scored a 90% score in diagnosing a medical condition from a case report (and explaining its reasoning) compared to just 74% composite score by skilled clinicians.
AI enthusiasts clearly have been informing health leaders that times are changing. As one explained, “Our individual health is heavily influenced by our lifestyle, nutrition, our environment, and access to care. These behavioral and social determinants and other endogenous and exogenous factors can now be tracked and measured by wearables and a range of medical devices.”
Information scientists remind us that both endogenous and exogenous information derived from our bodies can now be easily and continuously captured. It is estimated that 60% of health is the result of choice driven social determinants; 30% derived from genetic determinants; and 10% embedded in our past medical history. This suggests that both lifespan and healthspan are in part manageable. What remains an open question, for both citizens, and those who care for them, is who controls the data, and how is it protected?
As we’ve seen in prior sessions, the health care sector, which collectively consumes over $4 trillion in resources a year in the U.S., continues to balance profitability, productivity, and efficiency. It is not surprising then that institutional interest in AI is now focusing not only on data monitoring and diagnostics, but also customized therapeutics, operational technology, surgical tools, and clinical research trials. The number of sub-contractors involved in these efforts is enormous.
But as Mass General surgeon Jennifer Eckoff recently commented, “Not surprisingly, the technology’s biggest impact has been in the diagnostic specialties such as radiology, pathology, and dermatology.” Surgeon Danielle Walsh from University of Kentucky suggested that partnering was in everyone’s interest. “AI is not intended to replace radiologists. – it is there to help them find a needle in a haystack,” she said.
AI enabled technology is already embedded in many institutions. Consider AI-empowered multiphoton platforms for Pathologists. They say, “We anticipate a marked enhancement in both the breadth and depth of artificial intelligence applications within this field.” Experts already admit that the human mind alone can’t compete saying, ”Using nonlinear optical effects arising from the interaction of light with biological tissues, multiphoton microscopy (MPM) enables high-resolution label-free imaging of multiple intrinsic components across various human pathological tissues…By integrating the multiphoton atlas and diagnostic criteria into clinical workflow, AI-empowered multiphoton pathology promotes the establishment of a paradigm for multiphoton pathological diagnosis.”
One specific tool used for kidney biopsy specimens uses genAI. It was pre-trained through 100,000 inputs of the kidney’s microscopic cluster filters called glomeruli. Not only did this lead to superior diagnosis, but also a wealth of predictive data to guide individualized future care. There are of course “pro’s” and “con’s” whenever new technology pushes humans aside. The con’s here include the potential for human skills to be lost through non-use, less training of pathologists, decreases in intellectual debate, and poorer understanding of new findings. But these hardly outweigh the benefits including 24/7 availability, financial savings, equitable access, and standardization and accuracy.
Also new paradigms of care will surely appear. For example, pathologists are already asking, “Could we make a “tissue-less diagnosis simply with a external probe based on information markers emanating from a mass?” Diagnosis is also a beginning of a continuum that includes prognosis, clinical decision making, therapy, staging and grading, screening and future detection. What’s true in pathology is equally true in radiology. As one expert noted, “AI won’t replace radiologists. . .Radiologists who use AI will replace radiologists who don’t.” In one study of 38,44 mammogram images from 9,611 women, the AI system accurately predicted malignancy, and normal vs. abnormal scans 91% of the time compared to skilled human readings with 83% accuracy.
Our fourth AI and Medicine case is AI-assisted Surgery. Technology, tools, machines and equipment have long been a presence in modern day operating suites. Computers, Metaverse imaging, headlamps, laparoscopes, and operative microscopes are commonplace. But today’s AI-assisted surgical technology has moved aggressively into “decision-support.” Surgeon Christopher Tignanelli from the University of Minnesota says, “AI will analyze surgeries as they’re being done and potentially provide decision support to surgeons as they’re operating.”
The American College of Surgeons concurs: “By highlighting tools, monitoring operations, and sending alerts, AI-based surgical systems can map out an approach to each patient’s surgical needs and guide and streamline surgical procedures. AI is particularly effective in laparoscopic and robotic surgery, where a video screen can display information or guidance from AI during the operation.” Mass General’s Jennifer Eckoff goes a step further, “Based on its review of millions of surgical videos, AI has the ability to anticipate the next 15 to 30 seconds of an operation and provide additional oversight during the surgery.”
Surgical educators see enormous promise in AI-assisted education. One commented, “Most AI and robotic surgery experts seem to agree that the prospects of an AI-controlled surgical robot completely replacing human surgeons is improbable…but it will revolutionize nearly every area of the surgical profession.” Johnson and Johnson, a major manufacturer of AI surgical tools, had this to say, “Surgeons are a lot like high-performance athletes. New and learning surgeons want to see how they performed and learn from their performances and how others performed… Now, surgeons can look at what happened during procedures practically in real time and share the video with residents and peers, offering valuable post-case analysis and learning opportunities.”
By the way, in 2022 an AI Chatbot passed the US Medical Licensing Exam.
For surgeons everywhere, danger is lurking around every corner. It only takes a momentary lapse in concentration, an involuntary tremor or misstep to confront disaster. The desire for perfection will always fall short. But the new operative partnership between humans and AI driven machines, as in “Immersive Microsurgery,” reinforces accuracy and precision, is tremorless, and is informed by instructive real-time data derived from thousands of similar surgical cases in the past.
Case 5 in AI and Medicine focuses on “Equality and Equity.” Since 1980, the practice of medicine has relied heavily on clinical protocols. These decision trees, attached to a wide range of clinical symptoms and conditions, were designed to reinforce best practices across many locations with variable human and technologic resources.
No doubt, leaders in Medicine were caught off guard in July, 2022 when they opened the American Academy of Pediatrics (AAP) monthly journal and read the title of a formal Policy Statement: “Eliminating Race-Based Medicine.” The article stated, “Flawed science was used to solidify the permanence of race [and] reinforce the notions of racial superiority…. The roots of the false idea that race is a biological construct can be traced to efforts to draw distinctions between black and white people to justify slavery.”
The article tied American history to the problem of embedded bias, noting that the third US president, Thomas Jefferson, claimed in his 1781 treatise “Notes on the State of Virginia” that Black people had less kidney output, more heat tolerance, and poorer lung function than White individuals. The historic line drawn ended in a flawed protocol that had recently been exposed though AI-assisted investigation.
The “clinical protocols” or “clinical practice guidelines” had been constructed by proprietary medical education companies with the aid of leading academicians. Since they were proprietary, rationale and sourcing were not available. In one notable condition, following old AAP guidelines, Black children were under-treated in the Emergency Department for Urinary Tract Infection compared to White children with antibiotics. Once updated, treatment levels rose from 17% to 65% in Black children.
In another, algorithms for treatment of heart failure added three weighting points for non-blacks, assuring higher levels of therapeutic intervention for Whites. Lower rates of treatment of Blacks contributed to higher mortality. Women were especially vulnerable. Algorithms for success in Vaginal birth after prior C-section were scored lower if the mother was African American, assuring a higher level of repeat C-section in Black patients, and higher rates of postpartum complications. In summing up the situation, readers could not help but recall James Baldwin’s comment: “People are trapped in history and history is trapped in them.”
In the wake of these articles, clinical protocols across the board were subjected to AI-assisted scrubbing. AI was able to access massive databases, and turn an unemotional eye to the results. Generative AI is now in the lead under a new title – “Scalable Privilege.” The effort involves leveraging AI learning to utilize data from the best medical systems to better care for all.
This brings us full scale and keys up two essential questions when it comes to “AI and Medicine.”
- Are you prepared to accept massive health care system reforms in response to the unbiased findings that an AI driven assessment of population data will likely reveal? Stated another way, “Can we handle the truth?”
- What will you as a patient expect in return for granting unimpeded access to all of your de-identified medical data.
What we might as a nation learn from a population wide AI analysis could be quite disruptive and undermine the power and financing of multiple players in our system. AI would likely inform us that were we to follow its recommendations, health services would be more select, less biased overall, and less expensive. We would see fewer doctors, fewer drug ads, and fewer bills. But at the same time, that system might demand greater patience, greater personal responsibility and compliance with behavioral changes that ensure health.
Can we trust A.I.? That’s a question that AI Master Strategist Mark Minevich was recently asked. His response was, “There are no shortcuts to developing systems that earn enduring trust…transparency, accountability, and justice (must) govern exploration…as we forge tools to serve all people.”
What are those tools? He highlighted four: Risk Assessment; Regulatory Safeguards; Pragmatic Governance; and Public/Private Partnerships.
Like it or not, AI has arrived, and its’ impact on health systems in the U.S. will be substantial, disruptive, painful for some, but hopeful for many others.
Thank you, and the Presidents College at the University of Hartford for your participation.
