Marcus Weldon

By Marcus Weldon

Newsweek Contributing Editor for AI and President Emeritus of Bell Labs

0

Share

Newsweek is a Trust Project member

See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.

In a wide-ranging and illuminating conversation with the Dana-Farber Cancer Institute's exceptional duo of Drs. Ben Ebert and Eliezer Van Allen on the role they see for AI in medical practice, and oncology specifically, there was a moment when the dialogue stopped − in the way it does when you experience a moment of personal epiphany − and Van Allen uttered the phrase: "AI will re-humanize medicine." This summary framing captured everything that we had discussed but also embodied everything that a human (patient) feels about health care in the United States—that it is an inhuman experience, staffed by highly-qualified individuals with the best of intentions who endeavor to provide medical care via a process that is too often disempowering and dominated by excessive documentation and frequent disputes among the three P's of the medical industry: payers, providers and patients.

One can almost argue that the "analog" experience of health care that was prevalent 25 years ago has been lost in the pursuit of digitization of all data records and processes. At the recent Newsweek AI Impact Summit, Dr. Allen Chang, a practicing physician and Assistant Chief Medical Information Officer at the University of Massachusetts Memorial, opined that "Electronic health records [EHRs] have not measurably improved clinical outcomes." Indeed, there is widespread acceptance of the fact that they have not improved patient experience, clinicians' sense of job satisfaction, or even the operational efficiency of the entire health care value chain, which is now overrun with data requirements, requests and rejoinders.

This lamentable state of affairs is in stark contrast to goals outlined in a paper contemplating the potential roles of AI in health care published in 2021 by a team from Microsoft and the chair of medicine at University College, London, that argued that the "quadruple aim" of health care was "to improve population health, improve the patient's experience of care, enhance caregiver experience and reduce the rising cost of care." When measured against these clear goals, it is hard to argue that the digitization of health care has been anything other than a failure. But in the same paper, the authors conjecture — well before the current set of GenAI tools had their phenomenal progression — that "advances in AI have the potential to transform many aspects of health care, enabling a future that is more personalized, precise, predictive and portable."

Read More on Newsweek.AI

Against this backdrop, the conversation with doctors Ebert and Van Allen could not have been more timely or relevant. The combination of chief executive insight, medical practice experience and data science and AI expertise allowed a broad exploration of the entire health care and medical research space, with the clarity that comes only from being able to "embrace complexity to find simplicity" — arguably the most rare of intellectual skills.

The Re-humanization and Democratization of Health Care

...

One of the issues Ebert immediately focused on is how medicine is currently practiced. He agrees with David Eagleman's observation that language is a super-compressed package of knowledge — too compressed to convey all multisensory information and often specific to each person's subjective experience. Yet, this is how medicine is practiced today: "You have your patient visit, you get all of the information, and you summarize it in language, in a note, and then the next doctor reads that note and unpacks it. And that is the practice of medicine, and it's also the bane of doctors and patients," Ebert explained, adding, "you have to spend so much time either writing down or typing information to the computer" when there should be so much more to the practice and experience.

But this is one area where AI is already impacting medical practice with the increasingly widespread use of "ambient scribes" during doctors' patient visits. In short, an ambient scribe is the in-person equivalent of "meeting transcription" of the kind of video calls that became prevalent during the COVID-19 pandemic. Many voice-to-text specialist companies, such as Nuance (now owned by Microsoft) and Amazon's Healthscribe, offer services that allow a physical meeting between a patient and a doctor to be transcribed in real time, with auto-summarization into the EHR, allowing doctors to engage more directly with their patients, rather than their keyboards and computer screens. Ebert points out that this latter behavior was another unintended consequence of the move to EHRs, as complete records of every meeting became a requirement for the payer to reimburse the provider for services rendered and to authorize the requested treatments, follow-up actions or tests. By automating this process, clinicians have been able to save 4-6 minutes per patient, or 2-3 hours per day but, most importantly, both their job satisfaction and patient satisfaction increased manifold as they are able to re-establish a human connection with their patients and reduce or even eliminate their so-called "pajama time," the term doctors have come to use for the late night hours they put in to complete data entry requirements.

As Van Allen eloquently puts it: "AI holds the promise to actually make medicine more humanistic, to bring back what brought a lot of us into the field in the first place."

One of the most striking aspects of the conversation is one of the recurring themes of this series — Daniel Kahneman's "System 1 versus System 2" thinking paradigm — and the application to understanding the clinician's diagnostic process. Both doctors agreed that System 1-based heuristics dominate how clinicians diagnose, with facile symptom pattern recognition being the predominant method of diagnosis, with the more critical thinking System 2 mode frequently unengaged by the practitioner. That is due, in part, to the manifest reduction in critical thinking time that results from the data entry and management requirements in the EHR-mediated care process, and in part because that is invariably how we, as humans, operate by default: 95 percent of our brain's time is spent in System 1 mode and only 5 percent in System 2, in order to maximize response efficiency (in time, energy). In short, doctors are human.

However, Van Allen recalled how, when he was still in his residency, he and his fellow hospital clinicians would play a game based on House, the TV show starring Hugh Laurie as a master diagnostician, by attempting to unravel the unusual diagnosis that invariably anchored each episode before the big reveal on the show itself. This was a classic System 2 mode of exploration and synthesis that was highly stimulating but rarely practiced due to the aforementioned pressures inherent in current medical practice. Van Allen sees AI being an enabler of such critical thinking, acting as a creative thinker and challenger to the clinician, at speed, and with encyclopedia knowledge of the medical literature, both past and present — essentially providing a "House on demand." In many ways, this is similar the co-journeyist role that Rick Carter described in our interview but, in this "co-diagnostician" role, the AI system would have to be grounded in reality although permitted a higher "temperature" (the transformer model parameter, not the human metric in this case) in order to broadly search the massive symptom-cause-solution space.

This grounding clearly argues for the need for a biological or physiological "world model" that would describe the complex tapestry of biochemical interaction pathways, which could be used as the basis for intelligent exploration by AI. But Ebert cautions that invariably the rich interplay of pathways is often not well understood — we may understand a linear pathway for a given condition, but rarely do we understand the multitude of ancillary or complementary pathways that exist and form the essential basis for biological robustness. To quote Jeff Goldblum's character in Jurassic Park, Dr. Ian Malcolm, "nature always finds a way." For example, there are myriad so-called "escape pathways" employed by viruses, bacteria or cancers to effectively evade a host's immune response or a treatment regimen. This desirable evolutionary capability has the undesirable effect of making the deduction of successful treatment protocols a frustrating biological game of cat-and-mouse.

"One thing that I think that we are particularly bad at is we think that physiology and biology have simple homeostatic feedback mechanisms," Ebert explains. "You learn it as a linear pathway: A activates B, B activates C, C activates D, but it's not that simple. There are feedback loops all over the place that involve 20 other proteins and it's dynamic. And so, the pathway at rest is different from the pathway once it's activated and it's different from the pathway when it was activated two hours ago." The resulting complexities are staggering, Ebert says. "This happens all the time in cancer therapeutics, because you have that pathway you know that's activated in cancer and it's really important, and you get a drug that blocks protein C, so that should mean everything downstream of that protein D should be inactivated. But it doesn't work, because it turns out there are all kinds of clever [alternative] pathways and a bypass system activated D, even if C is inactivated. We don't comprehend those things well until we figure out the drug failed and why. Maybe that's partially a limitation in our knowledge, but it's partially the complexity of an actual pathway relative to the simplicity of the pathway that's in the textbook that you learn in college."

Human-AI model collaboration again seems to be the most productive or likely path forward. Van Allen highlights an example that illustrates the point: "A few years ago, we were trying to come up with a molecular signature in the DNA that would tell us whether a prostate cancer was lethal or not. And we had struggled to come up with a way to do this using traditional statistical approaches. And so, we said 'Let's try some AI.' And we thought, what if we imposed upon it what is known about biology and how we think signaling pathways work. How do we think the whole puzzle works together? There are open-source databases and ways that this is represented that we could borrow and insert into an otherwise fully connected neural network — a biologically informed neural network that represents the logic of what we know about biology. And we threw it at the same data we had been staring at for years, and it turned out it could actually do a better job than all of the other approaches that we had used up to that point in doing that task, and it was fully interpretable, and we could open up the box and start to query and understand and in this case, find a new drug target that might actually be useful for subsets of men with advanced prostate cancer."

But Van Allen then outlines the critical deficit we still face that more advanced AI models could assist with: "The problem with this approach is that fundamentally we are imposing upon it what we know about biology. And in the end, that roughly translates to about 4,000-5,000 genes of about 20,000 genes worth of information that we can map into any kind of structure. So, the representation we use is not only incomplete in terms of the entire genome's worth of information, but also likely incorrect. The future way of imagining how to do this would require taking much bigger biological data sets and applying some of the more advanced AI that now exists for language and other modalities into our domain. Then the challenge becomes how do we make this something queryable? How can we understand why [the model] is proposing this. I think these are things we're on the frontier of now, but we don't have answers yet. But that, I think, is where it's going."

Since the creation of complete biological foundation models for many conditions is not yet possible, in the near future we will have to rely on the written literature — a use case where LLMs are without doubt capable of superhuman capacity and recall. Both Ebert and Allen recognize the near-term potential of current LLM-based GenAI solutions and are unsurprised by the phenomenal results of recent diagnostic studies including one in which LLM-based systems were judged by a panel of expert clinicians to provide the correct diagnosis in 75 percent of cases, compared to around a 30 percent success rate for human clinicians. However, a human-in-the-loop remains essential, as evidenced by another study of systematic reviews of medical cases in which LLMs struggled with uncertain evidence and did not exhibit skepticism when studies possessed design flaws, resulting in these frontier LLMs failing to match the conclusions of human reviews in at least 37 percent of evaluated cases. In another recent study, clinicians working in tandem with an optimized clinical collaborator LLM (created using the GPT framework from OpenAI) achieved 90 percent correct diagnoses, more than 10% higher than using conventional (non-AI) reference sources.

A second area of real promise the two doctors clearly see is in drug discovery. The success of AlphaFold in predicting the folding structures of more than 20 million proteins, compared to the 170,000 protein structures that had previously been determined, is a testament to the potential for AI to massively accelerate the drug discovery process through such pioneering AI-based tools and methods. In addition, although the precise interaction pathways may not be fully understood, LLM-based search and discovery based on the existing literature has already shown significant success, with different studies reporting speed improvements ranging from 1.5X to 12X in pre-clinical research and development.

The role of AI goes beyond just intellectual conjecture; a team at Stanford has recently created an AI tool that will allow the automation of the process of biomedical research, stating that it "demonstrate[ed] robust generalization across diverse subfields, laying the groundwork for AI agents as integral collaborators in scientific discovery...with zero-shot performance across complex tasks — including those in genetics, genomics, microbiology, immunology, pharmacology, and clinical medicine." The authors note, however, that the system "struggles in areas requiring nuanced clinical judgment, novel experimental reasoning, analytical inventions, or deep biological thinking and synthesis. No system yet captures the full scope of human biomedical expertise." In combination, these observations succinctly capture the potential and the explicit need for AI and human (medical researcher) collaboration for optimized drug discovery and clinical trial management.

Looking beyond the research literature, the concept of drug-repurposing — in which interesting "side effects" were identified in clinical trials of a drug that were not related to the target application and so were not pursued further — is gaining significant momentum with the application of Gen AI approaches. Not-for-profit companies such as EveryCure, overall winner of the inaugural Newsweek AI Impact Awards, are using AI to analyze the wealth of clinical trial data to identify potential new therapeutics that already have regulatory approval for treatment of a given disease, and therefore would only require modification of the clinical practice guidelines for treatment of another disease. Understanding of the active pathway of action for the drug is often key to gaining this approval and, as described above, AI can again play a critical role.

Interestingly, Ebert and Van Allen also see other roles for AI in the clinical trial space. First, they highlight that matching potential participants to trials is a complex multidimensional factor analysis across thousands of trials and millions of patients. "Each trial around the country will have lists of inclusion criteria and exclusion criteria and no human brain can keep track of all of that. An AI algorithm could go through the patient's medical record and say that they meet the requirements. That's a huge advantage to patients," explains Dr. Ebert. It's also a huge advantage for researchers, he points out, because for trials to proceed at all they need to first enroll a minimum number of patients. Moreover, this analysis needs to occur not just at the point of initial diagnosis, but throughout the disease progression, as new trials emerge or eligibility requirements may be met later that weren't met initially. And Ebert sees this as another clear role for AI-based solutions. Indeed, McKinsey has recently reported the potential for a 12-month acceleration in clinical trial management using AI tools and techniques.

Our conversation then turned to the case that has been frequently reported in the media — that of the use of AI for radiological image interpretation. Prior to the rise of Gen AI tools, classical convolutional neural networks were used to read imaging data with increasingly high levels of accuracy. But even here, the right approach is to have a human in the loop. Ebert and Van Allen both cite examples where human knowledge is a vital complement to the AI-based image analysis. Ebert cites an example where the reported use of an AI for chest X-ray analysis led to the facile conclusion that radiologists were now superfluous. But on further analysis, it was found that the model did not work well for chest X-rays from a different hospital — for each hospital, there are subtle but important differences in the imaging employed and the human radiologist at each institution understands the specifics of their imaging technique. Any model would need to be trained on each hospital's data set and include the radiologist's knowledge. And, moreover, since the imaging quality changes and improves over time, the AI model would require continuous re-training to achieve the desired beyond-human performance.

Van Allen highlights another source of complexity: that the immunotherapy can't be understood based on information from the tumor alone, the drugs often work by acting at the interface between tumor cells and the immune cells that surround it. Ebert points to Hodgkin's disease as a prime example, as the number of malignant cells is small; the tumor mass is primarily made up of all the surrounding inert immune cells that are clumped together with their activity suppressed due to mutations in the tumor that block the immune system reaction. "This is an extreme example where there's so much of the surrounding cells involved," he says, "and that's actually the core biology of the whole process."

Van Allen generalizes this observation to pathological image analyses — citing the fact that there are also very subtle hospital-specific biases that are introduced in the preparation of the tissue slides which, when combined with varying ambient temperature of the lab and even the lighting conditions, make generic interpretation subject to significant error. He highlights the value of incorporating human "intuition" or practice into these models, citing an early example he experienced with a colleague working on an AI-based image recognition tool. "He realized that just blasting an AI without any domain knowledge ended up yielding something that didn't work very well and was uninterpretable. He was a PhD scientist, not a radiologist, so he went back and started talking to a bunch of radiologists who were experts in reading mammograms. He rebuilt the AI to effectively mimic what the radiologists were doing and then found that it not only worked as well or better than the radiologists [but could] do some of the different kinds of decision support as well."

In summary, both doctors see the undeniable potential for using AI tools in image analysis across the health care space, but also the need for the models to be trained on a comprehensive data set that accounts for all the imaging specifics of each lab and with knowledge of the specifics of each disease and incorporating human expertise and best practice.

This discussion of the need for a comprehensive data set that represents the space recalls the Klein-Kahneman definition of learnable systems, which have two fundamental requirements:

Van Allen and Ebert acknowledge that satisfying both criteria is challenging for many areas of the medical field, due to the scope of the variations, the complexities of the systems and the disjoint or incoherent nature of the data sets. So, until these criteria can be satisfied for a given case, a supervising specialist should always be present in the loop and retain the final say, a conclusion that is consistent throughout our conversation.

Beyond the clear opportunities discussed above in doctor-patient interaction, drug discovery, clinical diagnosis (including image analysis) and clinical trial management, Van Allen and Ebert see a multitude of other opportunities for AI in the health care field:

A recent study of the primary state-of-the-art LLMs supports this optimistic view: on an accuracy scale zero to one, most models were found to perform strongly in; clinical note generation (score: 0.74-0.85); clinical decision support (score: 0.61-0.76); patient communication & education (score: 0.76-0.89); medical research assistance (score: 0.65-0.75) and; administration & workflow management (score: 0.53-0.63).

Looking across this multiplicity of AI benefits, it is clear that van Ellen could not have summarized it better: taken together, they will rehumanize health care and manifestly achieve the "quadruple aims," by making health care more "personalized, precise, predictive and portable," as imagined in the 2021 paper. And part of this re-humanization includes the democratization of health care that will allow care to be provided anywhere, anytime, using the available human providers and equipment, augmented by AI, allowing operation to the absolute limit of the provider's abilities.

The two doctors recognize that medicine is a conservative discipline, so progress is likely to be "methodical," but the best way forward is to have clinicians and scientists/researchers working hand-in-hand to create optimal AI models. They also have an ace up their sleeve: Dana-Farber is in the midst of planning and building a new facility in Boston that promises to be "the first fully AI-enabled hospital." When asked what that would enable from a patient experience, the two could barely contain themselves with a torrent of ideas and possibilities. Dr. Van Allen painted a picture of the future he sees:

"It might be an hour before the nurse, or a doctor goes in, and you could monitor that in real time and actually have the data analyzed. There are physiologic signs that were happening in a person's breathing pattern or their temperature, and the patient didn't even realize that their heart rate's starting to go up, their breathing starting to change. And the next time they'll get their vitals might not be for some period of time. But you can build into the hospital infrastructure the capacity to measure these patients in ways that are currently not possible, or not even remotely conceivable, but we have the chance to build something from scratch. And then on the flip side, imagine a scenario where that hospital has all of the sophisticated technologies to generate the data and create personalized therapeutics that are built on top of foundational AI technologies around drug development, drug design, cell therapies and so on and so forth, that are maximizing information about the patient's molecular tumor information and the drug development capabilities. So, in that hospital setting, they're infusing the most personalized and conceivably most effective drug for that patient at the right time. And then, lastly, all of this information can be learned and gathered and to help build the next cancer foundation model, so that we really get towards the dream of that high-dimensional-impossible-for-a-human-to-wrap-their-head-around space of information that can yield the next big discoveries, the next big clinical insights and the next big ways that can change care everywhere."

That's a health care dream — soon to be a reality — that's hard to beat.

View 0 Comments

Request Reprint & Licensing

Submit Correction

View Editorial & AI Guidelines