Lessons from Brian Anderson, CEO and Co-Founder of CHAI and Chief Digital Health Physician MITRE, on building trustworthy AI that serves all of us
Brian Anderson, CEO and Co-Founder of CHAI, a community of academic health systems, organizations, and expert practitioners of artificial intelligence and data science
Subscribe to our substack for updates and listen on Apple Podcasts or Spotify.
Welcome back to the Pear Healthcare Playbook! Every week, we’ll be getting to know trailblazing healthcare leaders and dive into building a digital health business from 0 to 1.
This week, we’re super excited to chat with Brian Anderson, CEO and Co-founder of CHAI (Coalition for Health AI), and Chief Digital Health Physician at MITRE.
This week, we’re super excited to chat with Brian Anderson, co-founder of CHAI, Coalition for Health AI, and Chief Digital Health Physician at MITRE.
CHAI is a community of academic health systems, organizations, and expert practitioners of artificial intelligence and data science. Brian leads research and development efforts across major strategic initiatives in digital health, partnering with the US government and private sector companies.
Prior to joining MITRE, Anderson led the Informatics and Network Medicine Division at athenahealth, where he launched a new model of clinical decision support leveraging artificial intelligence. He has also served on several national health information technology committees in partnership with the Office of the National Coordinator (ONC). On top of all that, Brian is also a Harvard Medical School Associate Professor of Biomedical Informatics and senior advisor for ARPA-H in Clinical Trial Innovation. Anderson completed his clinical training at Massachusetts General Hospital and completed his B.A. and M.D. degrees at Harvard College and Harvard Medical School.
In this episode, we talk about Brian’s career journey from clinical practice to digital health, CHAI’s efforts to establish AI process standards for healthcare, the concept of a federated network of assurance labs, and advice for the health AI innovation community.
Brian’s journey from clinical practice to Athena:
While practicing at Massachusetts General Hospital, Brian helped install a new EHR that was coming online in the clinic. It was challenging to use:
“My eyes were really opened to the impact that a poorly configured EHR can have on a physician's lifestyle, a physician’s work-life balance. Mine took a turn for the worse.”
Brian wasn’t getting home until 8:30 at night because of the EHR’s low usability. He first stumbled upon Athena as a clinician looking for a cloud-based EHR solution but, on the advice of a friend, ended up taking an interview with Athena’s Chief Medical Officer. Brian was inspired to jump into the deep end of the clinical EHR division.
At Athena, Brian worked on rethinking EHRs to enable providers to tell their story in partnership with patients. Brian shares that EHR workflows weren’t configured to work with physicians until around the early 2000s and 2010s; EHRs started off as revenue cycle management mainly focused on billing optimization. The industry eventually became more sensitive to physician burnout, searching for a way to use digital tooling to enable a more connected physician-patient relationship.
“We didn’t get into medicine to stare at computer screens and do documentation faster, we got into medicine because we wanted to connect with our patients and make a difference in their lives. Having technology now that enables that is really exciting. It’s empowering.”
MITRE is a not-for-profit corporation committed to the public interest, operating federally funded R&D centers on behalf of U.S. government sponsors. They don’t sell anything commercially, making it easier to work with private sector partners given the guarantee that MITRE will not sell your data.
CHAI (Coalition for Health AI) is a community of regulators and innovators— academic health systems, organizations, and expert practitioners of AI and data science.
Brian joined MITRE in 2018. Coming out of the pandemic, Brian and his co-workers debated the question: how do we work together better? How do we build trust in digital technology that doesn’t reinforce the digital divide, but rather enables more ethical, meaningful tools?
“AI is obviously very impactful. There’s a real need to have a set of guidelines, shared principles that regulators are going to want to refer to, that our government wants to safeguard us as citizens against the kinds of AI that could hurt us.”
Brian shares that when they looked at the health AI space, it was clear that there wasn’t a codified set of standards yet. The team launched CHAI, Coalition for Health AI, with a vision of generating AI process standards.
The group started with a half-dozen organizations including health systems like Mayo and big tech companies like Microsoft. During the first year, CHAI developed a blueprint for trustworthy AI that was recently released. Brian shares that their goal was to create a target for innovators in this space: what is our shared goal for what responsible AI looks like?
CHAI soon grew from a half-dozen organizations to 200 organizations, then 300, then 700 organizations. The US government caught wind and everyone from the White House to various HHS agencies joined.
Brian shares that the key to representing different stakeholders involves 1) aligned incentives and 2) identification of a clear pre-competitive space that enables individuals to have successful business models. The goal of their framework is to allow groups to protect IP but not slow progress down.
Advice for the private sector innovation community
First off— how can startups get involved with CHAI? Why should they?
CHAI is a 501(c)(6), a membership-driven organization mandated to have representation in its governance and operational process. Brian shares that CHAI now has over 1,500 different organizations, many of which are startups, and everyone in the group contributes to the standards that will be published.
“The real opportunity is to be part of building the definition of what those standards are, and what that testing and evaluation framework is, from within these working groups.
The challenge is startups don't have a whole team for responsible AI like a big tech company might. You need to be really thoughtful in where you put resources, because joining these working groups is a resource ask.”
Brian shares that startups have an agility advantage— they can quickly implement these standards to give feedback to the community on what’s working and what’s not. “The startup community can really be in the driver’s seat in terms of helping to drive that interactive feedback process.”
The Department of Veteran Affairs has just launched their AI tech sprint, one example of an opportunity that allows startups to participate among big tech giants.
Understand what the target looks like from the regulatory perspective.
Brian debunks one of the misunderstandings about government in this space— he argues that the government doesn’t want to mandate technical standards if it’ll stifle innovation. Their goal is to build a regulatory framework after there’s a shared target of what good AI looks like. Additionally, Brian highlights that this movement started not with government regulators, but innovators coming together to define a consensus opinion.
The regulatory community will not only be referring to success frameworks, but also testing and evaluation frameworks as part of their decision making process. Brian shares that CHAI is now working on creating consensus on testing metrics and methodology, measuring performance, measuring bias in generative AI, etc.
Validation at every stage is important.
Brian shares that AI assurance tools will be created for all phases through a model’s life cycle: development, deployment, monitoring, and governance. It’s easier to mitigate bias issues iteratively in the development of a model using tools that can accurately help you understand the model’s performance from the earliest stage.
He encourages entrepreneurs in the development phase to think about using these tools while determining what the models should do.
“Think with equity in mind from the beginning.”
Answer these critical questions: What is my model’s economic benefit? Is my model trustworthy? Bring transparency, explainability, and interpretability to providers.
Health systems are the majority of the customers of these models, and Brian shares their biggest question: what is the economic benefit of using this model? The most successful entrepreneurs in this space make the economic problem they’re trying to solve very clear.
“It can be as straightforward as clinical efficiency, helping providers see more patients. There's a lot of ambient technology, voice-to-text solutions in that space that are doing quite well, that show real ROI.
Health systems are really challenged. A lot of them are in the red. The idea of spending millions of dollars on something that might or might not have an ROI is really challenging for them.”
The other essential quality Brian highlights is building trustworthy AI. Clinicians are going to be playing a role in a big health system’s decision to move forward with your model or not. Brian encourages the innovation community to focus on explaining the output from these models and help providers answer the critical question: is this model fit for purpose given the specific use case or patient I have?
Key takeaways from CHAI’s Blueprint on AI
“It’s one thing for a vendor that develops models to say ‘trust us, our model performs.’ It’s altogether another thing for a customer to say ‘that’s great, but I want that model to be independently validated.’
If you think about purchasing a car— if you’re very thorough and do your research, you often look at the car test reports from the National Highway Traffic Safety Administration or the car insurance. That’s what we envision these assurance labs to be able to support.”
Brian’s vision for a federated network of assurance labs: building AI that serves all of us.
Generate diverse datasets. A federated network of labs would create diverse, heterogeneous sets of testing data that are representative of different segments of our population. “If the goal is to build AI that serves all of us, you can’t have one testing lab. You need them all over the place.”
Focusing on usability: Brian believes that assurance labs won’t have a metric-centric approach alone. As a part of the validation process, they’ll also be looking at how the model is configured in the EHR and the clinician workflow.
“Workflow is profoundly impactful and influential, potentially creating bias that we want to be able to account for.”
Accelerate training. Beyond testing and validation, data from a network of 50+ labs in the US could get us closer to what Brian calls a “learning healthcare system”.
Brian shares that CHAI’s goal is to empower organizations beyond the nation’s most preeminent academic medical centers — FQHCs, critical access hospitals, rural clinics — with AI assurance toolkits so that they too can do validation. These centers often have patients that AMCs (academic medical centers) don’t.
For more information and utilities regarding the Assurance Labs you can read more about it here in JAMA.
Brian highlights Mayo Clinic’s Platform_Validate as one of the first assurance labs that’s been launched and shares that about 30 other health systems are in the process of standing one up.
He predicts that health systems that are on common cloud platforms like Azure, GCP, and AWS will likely partner with their platform vendors to build AI assurance tools.
“How do we build that kind of insight and equity across the ecosystem of digital health? It’s really fascinating to see that AI may be the thing that finally gets us over the finish line to realizing what a learning healthcare system could be.”
Brian’s predictions for the future of healthcare AI
Brian believes it’s helpful to look towards AI usage in other industries like aerospace and critical infrastructure. The approach is the same across the board: what is the risk of the technology failing? What are the consequences if it fails? Looking at other industries, Brian sees automation that has pushed the risk curve to a more and more acceptable space. He believes healthcare will get there too, once the technology is safer.
In the spectrum of assisting, augmenting, and automating clinical work, Brian believes we’re squarely in the space of developing solutions for augmentation and assisting clinicians— automation is on the way.
Before we can get there, Brian believes we need to have answers to these questions: How do we build models with greater equity in mind? How do we define and agree upon a method of measuring performance and accuracy?
At this point, Brian shares that we don’t know how to measure performance in the generative AI space. We don’t have an agreed upon approach, there are emergent capabilities that we can’t predict.
“If all algorithms are programs trained on our histories, our history is imperfect. The question then is, how do we train algorithms in a way that builds a better future? We don’t necessarily have an agreed upon way to do that.”
We’re already beginning to see that models are better at differential diagnosis, and that’s a sign of the automated spaces that are to come.
Within health AI, Brian’s excited about two main areas: access and navigation.
Brian’s excited about personalized patient navigation tools that can take in and ingest a lot of data— taking in your own personal health record to navigate a complex health ecosystem, getting the highest quality care at a lower cost.
Drug discovery. Brian shares that he spends a lot of time in this space, and they’ve seen the regular advancement in the capabilities of these models.
LCMs (Large Chemical Models) are able to accurately predict not only molecular composition but also the protein fold, the function of the molecule, and more— all in silico.
The in vitro space is also advancing quickly. Building personalized tailored medication, assisted with AI, could help bring drugs to market cheaper and faster.
Last year, Brian would have said that we were 10 years away from AGI (Artificial General Intelligence). Now, he estimates we’re closer to 2-5 years away.
“We’re at a moment where we can choose to reinforce the digital divide. If all algorithms are programs trained on our histories, there’s a whole subset of people that haven’t been able to tell their digital history or had systemic injustices reinforce their history.
AI gives us the opportunity to ask ourselves the profound question: do we want to change that? We have a way to do that. All we have to do is identify the data to train the model on so that it can be more equitable, it can be more fair.”
Interested in learning more about MITRE or CHAI? Learn more on their MITRE website, CHAI website, X, and LinkedIn.