Chairman Guthrie, Ranking Member Eshoo, Chairwoman McMorris Rodgers, Ranking
Member Pallone, and distinguished members of the Committee, it is my pleasure to appear before you today to discuss how artificial intelligence is changing
healthcare.
My name is Dr. Benjamin Nguyen. I am a Senior Product Manager at Transcarent,
leading our AI team, which is tasked with improving the care experience, the quality
of care, and the affordability of care by expanding Transcarent’s affiliated virtual
clinic’s suite of AI tools while maintaining the highest standards for patient safety.
By background, I am a medical doctor and completed my degree at the Keck
Medical School at the University of Southern California. I have worked at the
intersection of technology and care delivery throughout my career, with a special
focus on artificial intelligence.
About Transcarent
Transcarent was founded to make it easy for people to access high quality,
affordable care and offer greater choice and control over healthcare for those that
pay for care, including healthcare consumers (our Members) and employer-
sponsored group health plans. Specifically, we offer access to a physician or
caregiver via chat in 2 minutes or less, a pharmacy marketplace that allows people
to find the lower cost for drugs and the locations, care provided at home or close to
it for more affordable prices, surgery information on necessity, quality, and cost,
and care guidance for complex conditions like cancer. We also are working on
specialty medication programs to make GLP-1s both more affordable and more
value-based.
Transcarent is not a health plan. Rather, services offered through Transcarent make
the Member’s journey more informed, easier, and more affordable and we help to
make their existing medical plan easier to understand and use.
Transcarent cuts through the complexity of the current healthcare system, making
it easier for people to access high-quality, affordable care. Our platform is
personalized for each Member, offers access to physicians, nurses, therapists and
other healthcare professionals on-demand. With a connected ecosystem of
high-quality, in-person care and virtual point solutions, Transcarent confidently guides Members to the right level of care.
Transcarent Members have access to care through digital guidance along with the
clinicians in Transcarent’s affiliated virtual clinic as well as from high-quality
providers in their local communities. We provide one place for all their health and
care needs. Transcarent Members access complex care resources for behavioral
health, musculoskeletal or oncology care virtually or in-person, often in value-based
care arrangements that align incentives between those providing care and those
paying for care. We use the power of technology to scale access to high-quality
care, regardless of a person’s geography, income, education, ethnicity, disabilities,
gender, or language.
Today I will speak to the applications of AI within our affiliated virtual clinic. At
Transcarent, we believe AI holds great promise to revolutionize care delivery for the
better – to speed access to treatment, reduce administrative burdens on clinicians,
and democratize access to healthcare resources (for example, it no longer matters
where you live if you want to talk with an endocrinologist, whether in a rural area or
a major metropolitan city.) We also believe that AI should be used only for the right reasons – it should not be used to deny access to appropriate, medically necessary
care or in a way that diminishes the quality or safety of that care.
Scope of This Statement
AI is an extremely broad term, used inconsistently even by those who work in the
industry. There are many valuable types of AI, such as decision support systems or
voice recognition AI systems. I chose to focus this statement on a specific, new kind
of AI known as Generative AI--the underlying technology powering products like
ChatGPT. I chose to focus on Generative AI systems because they have advanced at
an unimaginable velocity in recent years and can have a significant positive impact
on the healthcare industry.
AI Technology Has Changed
To understand the importance of the recommendations made in this statement, it is
important to understand recent changes in AI technology. It is common to think
about the advancement of technology along a smooth continuum, like a train track
going up a hill. But sometimes, there are technological leaps that are so great that
they propel us rapidly forward into a different world. A very large leap occurred in
the AI industry over the last few years, specifically in a subtype of Generative AI
technology known as generative large language models (LLMs). The magnitude of
this change is so large that it is like going from locomotives to powered flight. Like
the advancement from locomotives to planes, this leap brings new opportunities
and also risks. It is crucial that Members of Congress understand what has changed
because the nature of the changes will drive the nature of the opportunities and
risks in American healthcare.
Generative LLMs can be used in many domains, but they are most well-known for
being the engine behind next-generation chatbots, like ChatGPT. As AI engines,
Generative LLMs are computing models that can parse and generate fluent written
language. Three key characteristics distinguish Generative LLMs from the AI that
most people are used to seeing.
First, chatbots that are powered by Generative LLMs are surprisingly good at
handling complex topics, understanding nuanced contexts, and weaving them into
their decision-making and answers. This means that unlike chatbots built on older AI
technology, chatbots built on Generative LLMs can adapt to their users’ complex
requests and needs, holding complex and nuanced extended conversations, and
handling tasks that require logical reasoning. These abilities set these new AI
models apart from previous generations, enabling them to engage in tasks that AI
systems of the past could not.
Second, Generative LLMs are by nature, very flexible. This contrasts with older
“narrow” AI systems. Older, Narrow AI systems are built in such a way that they can
only function within a narrow domain. If I build a Narrow AI system that answers
basic questions about the branches of the government, I cannot ask it to generate a
poem in the style of Shakespeare or to give me career advice unless I rebuild the AI
model from scratch. On the other hand, a single modern Generative LLM given
sufficient inputs can handle any of these tasks easily, despite not having been
explicitly designed for any of them. Some Generative AI products also combine
powerful image recognition AI systems with Generative LLMs, enabling them to
interpret scanned documents or even images and other media and incorporate the
context into their answers.
Third, and perhaps the most important distinguishing factor, is that the full
capabilities and limitations of these new AI systems are not fully known, even by
their developers. This feature of Generative AI means that bias is even more
challenging to detect when compared to older Narrow AI systems. There are many
reasons for this. First, as with most modern AI systems, Generative LLMs must be
“trained” on massive volumes of written language – the ultimate compendium of
human experience. It therefore inherits the inherent biases of that experience
through the data used to train the model. Second, during their development phase,
most Generative LLMs depend heavily on a large number of human workers who
“grade” their performance, helping to “teach” the AI system to hone its capabilities.
Bias can be introduced in this stage inadvertently, both through bias introduced by
the selection of human workers and through the subconscious biases of these
workers themselves. Third, Generative LLMs, especially when used in chatbots, can
accept open-ended written inputs and create open-ended written outputs. Compare
this to a simple Narrow AI system for predicting whether it rains or not. We know
that the worst this Narrow AI system can do is mispredict the weather because it is
a simpler machine with fewer parts that can break. On the other hand, a Generative
LLM chatbot can take any written input, and it may respond in countless ways
based on its internal probabilities. I might have no idea that it can be biased until I
ask it the right questions, using the right combination of words. When the ways that
an AI system can be used multiply, so do the potential risks.
Opportunities in Healthcare
In healthcare, some tasks are difficult because they require deep expertise and
clinical judgment, and there are tasks that are difficult because they are tedious
and labor-intensive, yet necessary. Generative LLMs will someday be applied to the
former. But today, they can already excel at the latter. And that’s good news
because for patients (which we call health consumers), administrators, clinicians,
and the multitude of supporting staff charged with delivering care to people, those
tedious and labor-intensive tasks are the work that makes our healthcare industry
tick. For instance, it has been estimated that 30% of healthcare costs are
administrative.1 With proper design and testing, Generative LLMs are very well
suited to the everyday administrative work of information synthesis, documentation
creation, and form filling which can reduce administrative burden and staff burnout.
These are obvious use cases, and they will be Generative AI’s first proving ground in
healthcare.
But these uses, as incredibly valuable as they are, will be relatively incremental in
terms of impact in comparison to coming applications in medical triage, navigation,
and personalization. The great leaps will come from transforming the way that
everyday Americans relate to their healthcare system. The American healthcare
system is complicated and confusing. To get healthcare, the average person exerts
immense effort. They wade through the complexity of insurance coverage, copays,
deductibles, in- and out-of-network clinicians and benefits, doctors' schedules, and
office locations – all before getting care. It is a wonder that patients even have the
energy left over to make sound, well-informed medical decisions. Generative AI is a
major step towards simplifying the healthcare system so patients can focus on their
health and the health of their loved ones. A well-designed Generative AI health
navigation product would allow a person to simply state that he or she needs an
appointment with a doctor – and an appointment will be made with an in-network
doctor who speaks the language he is most comfortable with, who has a time
available during the patient’s lunch break and is within a 10-minute drive. The
person can and should only need to focus on getting healthy.
The complexity and dynamics of healthcare also lead to another kind of inequity -
“one size fits all.” As the country moves to a more value-based care approach, we
need to deliver more personalized, high-impact care. However, different people
have different needs. Some want to ask their doctor more questions than others.
Some have a harder time understanding the side effects of their medications.
Others need help deciphering lifestyle modifications to stave off heart disease. Yet
we are all bound by the typical 10–15-minute office visit, and we receive the same
post-visit brochures, written with the same talking points. The fact that most
Americans rely heavily on the internet to supplement their understanding reflects
our failure to meet these diverse needs.2 A Generative AI product could create
educational materials that are tailored to the patient’s level of health literacy and
education, in the patient’s preferred language, including content that addresses the
actual questions that the patient asked during their visit, helping to relieve their
most salient healthcare concerns. A Generative AI chatbot working hand in hand
with a doctor can enable a medical visit to go for as long as it needs to, by
conversing with the patient, answering all their questions fluently in the language of
their choosing, and allowing them to ask for clarification or simplification. This is
one way that Generative AI can reduce the structural biases of the healthcare
system – by moving us from “one size fits all” to “many sizes for many needs.”
Other impactful uses of Generative AI in healthcare include but are not limited to:
personalized behavioral modification plans to reduce chronic disease risk,
considering specific patient preferences, budget and geographic constraints,
and family support;assisting patients to apply behavioral health coping strategies in conjunction
with medication and talk therapy;organizing and synthesizing research and expert information for healthcare
practitioners and even patients and families;training and education for medical professionals, personalized to individual
needs and levels of expertise; andguiding and supporting families and patients through complex medical
decisions.
How Transcarent Uses AI
When a person comes to Transcarent’s affiliated virtual Clinic, an AI assistant
immediately begins to gather information from them about the reason for their visit,
organizing it for the clinician. By the time the clinician greets the individual, they
have a detailed and relevant summary of their symptoms and history. This reduces
administrative burden and allows the clinician to spend their time focused on
diagnosis, treatment decisions, and working in partnership with the person on
follow-up or preventive care. This approach serves both clinicians and the people for
whom they are caring.
Perspectives for the Future
My own journey from medical education to building AI products has given me some
perspective that I hope will be valuable to the members of this Subcommittee.
First – norms of responsible AI healthcare product design must be established
deliberately and thoughtfully. This does not necessarily mean new regulation – but
it does mean that stakeholders across the healthcare industry must come together
to share knowledge and establish what responsible AI use in healthcare means. The
AI industry’s biggest players have established their own internal AI safety divisions
due to the potential harm that AI can cause. Applying AI in a unique and
multifaceted industry like healthcare will create unique and multifaceted ways that
it can cause harm. For example, we may want to establish the principle that AI
should never be used to make decisions to deny appropriate medically necessary
care.
In the absence of healthcare-specific frameworks, we have established a proposed
set of core principles to govern our AI product development at Transcarent:
Patient safety - Patient safety means that any AI products we build must
include redundant safety mechanisms as a core part of their design – safety
systems surrounding AI are our “table stakes” features, not afterthoughts.
We build AI products to augment and enable clinician decision-making and
diagnoses. Our AI products are not diagnosing people --- that responsibility
remains with the clinician.AI equity end-to-end - AI equity end-to-end means that we consider how
decisions along the entire product development process will impact equity.
Will a data set used in development inject bias into our AI product? Or will
applying AI technology in a different domain than was originally intended
lead to biased results?Patient- and clinician-centered AI – We design our AI products to enhance
the experience of our patients and clinicians. That means that we consider
them key stakeholders during the design process, and work to ensure that
our AI products earn their trust and are built to serve their needs first and
foremost.
Second, there is a significant and growing gap in AI talent within the healthcare
industry’s practitioners and leaders. Healthcare’s unique challenges and
opportunities mean that we need to develop internal expertise in AI and Generative
AI. Even amongst AI companies and experts, there is not good agreement yet on
how to even measure the capabilities and safety risks of new AI technologies, much
less how to mitigate the risks. The development of healthcare AI products that serve
all Americans equitably demands active participation from all levels of the
healthcare system, from executives to those providing care to patients at the
bedside. Generative AI systems are so complex that the study of their biases,
capabilities, and potential for malfeasance or beneficence to society has become a
discipline unto itself (known as “alignment research”). The researchers who study
Generative AI are often more akin to biologists running experiments than engineers
building machines.
Thus, we must also study Generative AI as a discipline within healthcare, or we risk
leaping into the future without understanding the consequences unique to our
industry. Having gone to medical school, I know that the training of individuals in a
discipline orthogonal to the hard medical sciences does not come naturally to our
institutions. Most medical students and doctors do not have pathways, nor do
medical institutions have the funding or expertise to create them. We need
incentives, frameworks, and funding to create these pathways if we want to be able
to ensure Generative AI can achieve its potential to revolutionize our healthcare
system.
In conclusion, I believe that the integration of Generative AI into healthcare, as with many other industries, holds great promise. With costs skyrocketing and care becoming less accessible to the average American, AI-enabled platforms, like Transcarent, will help make high-quality, affordable care accessible to more people who need it. At the same time, it is critical to balance the promise of AI with safeguards that will be necessary to build and preserve trust across all stakeholders.
Thank you for your attention and I am available to address any questions you may
have.
1 High U.S. health care spending: Where is it all going? (2023). Commonwealth Fund.
2 Finney Rutten LJ, Blake KD, Greenberg-Worisek AJ, Allen SV, Moser RP, Hesse BW. Online Health Information Seeking Among US Adults: Measuring Progress Toward a Healthy People 2020 Objective. Public Health Reports. 2019;134(6):617-625. doi:10.1177/0033354919874074
