Lessons from Dr. Matthew Lungren, Chief Data Science Officer, Microsoft, on the impact of AI research on healthcare
Dr. Matthew Lungren, Chief Data Science Officer of Microsoft Life Sciences and Healthcare and Adjunct Professor at Stanford Department of Biomedical 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 Dr. Matthew Lungren, Chief Data Science Officer for Microsoft Health & Life Sciences
At Microsoft, Dr. Lungren focuses on translating cutting-edge technology, including general AI and cloud services, into innovative healthcare applications. As a physician and clinical machine learning researcher, he maintains a part-time clinical practice at UCSF while also continuing his research and teaching roles as adjunct professor at Stanford.
Prior to his role at Microsoft, Dr. Lungren was a clinical interventional radiologist and research faculty at Stanford Medical School where he led the Stanford Center for artificial intelligence and medicine and imaging. His scientific work has led to more than 150 publications including work on multimodal data fusion models for healthcare applications. On top of all of that, Dr. Lungren is also a top-rated instructor on Coursera with his AI and Healthcare course designed to help learners with non-technical backgrounds.
It was really fun to chat with Dr Lungren about his clinical, research and private sector experiences as a Professor and physician at Stanford and UCSF to leading Microsoft HLS and Nuance. We discussed how technology within the radiology field has developed throughout time, and exciting new Microsoft HLS Research projects and the impact of AI research on healthcare. We also had rapid fire fun facts at the end! Hope you enjoy it!
Listen here for the full content:
Dr. Lungren’s clinical journey as a radiologist:
Matt was an English and biology double major as an undergrad, disciplines that people perceived were completely different. He shares that he’s always been drawn to the intersection of disciplines:
“When you boil it down, we’re getting at the same core truth: what is the human experience? How do we learn about diverse perspectives to enrich our understanding and solve problems?”
Matt has continued to be drawn to intersections, today, to AI and science and how the two merge together.
As a clinician, he was attracted to radiology as a speciality because of the technology and because of the opportunity to interact with every specialty across the health system. Matt shares that he appreciates how multidisciplinary radiology is— the ability to tie things together in really interesting ways, particularly in interventional radiology. Those that started the field were incredibly creative, coming up with the stent, the angioplasty, inventing things on the fly. There wasn’t a dogma that dictated what you had to do, and the discipline encouraged solving complex problems in innovative ways.
The evolution of machine learning in radiology: Before computer vision took off, the building process looked different. In the days of building handcrafted features, you would ask experts to come up with their decision process and try to operationalize that with machine learning.
Over time, Matt has seen Richard Sutton’s “Bitter Lesson” play out: scale dwarfs building handcrafted rules, allowing models to learn. Computer vision kicked off ImageNet, which was a huge breakthrough. People in the medical space began to imagine how it would work for healthcare.
Matt loved the curious nature of his field, and people recognized his potential to apply that to research. He got a call from Sam Gambhir, the former chair of radiology at Stanford, encouraging him to pursue the research side of medicine. Matt got hooked on digital health when he saw how imaging data could be used to find population health insights.
“I think one of the guiding principles early on, which wasn't very common in the digital health space, was: we're going to make as much data as we can public. That was not an easy battle to fight, but it has paid off in ways that I can't even list. It's been incredible to see the innovation that's come from the wisdom of the community.”
As Chief Data Science Officer of Microsoft Health & Life Sciences, Matt’s role sits across the organization. There’s an HLS group, but he also works with Microsoft Research to convert research into roadmaps.
Microsoft’s vision: “Empower everyone to achieve more.”
After his stint in academic medicine, Matt saw an opportunity to translate his work in an industry role. He knew that communication between tech and medical stakeholders was never perfect and saw an opportunity to “get a lot more done”. When Nuance was acquired by Microsoft, change began to happen at more scale. In this new era of AI, the timescales for technology have sped up even more. Matt cites DAX Copilot as one example that’s reducing administrative burdens within EHRs.
“There’s certainly easier sectors of the economy to play in… but there’s something fundamentally meaningful about it.
You can apply the talents that you have to anything. You can make search .001% better, you can serve an ad .001% better, but you can also make someone’s life better. I know that that sounds like a sales pitch. But, ultimately there is this sense of meaning that you derive from working in healthcare no matter what role you're playing in it.”
Matt used to choose projects based on timeline— it would take many months to gather data, build a solution, and at the end of the day, only one thing would be solved. The timelines have collapsed. He was surprised by how much a general model can accomplish. Now, he can pick use cases that weren’t at the top of the list, but still solve problems, leveraging that general technology.
“It’s completely transformed how I think about what to take on and what’s important to prioritize.”
On Microsoft HLS upcoming projects:
Multi-modal is in rapid development. Microsoft Research has already published some papers on this, one interesting area being RAD-DINO.
DINOv2 was a family of models from Meta that allowed great embeddings from imaging data. Matt shares that his team began to rethink using joint representation to combine these modalities. They realized that the classic technique of using image-to-text joint representation wasn’t the best way to do this within medicine.
If the only output from images is what a radiologist says about it, what do you leave on the table? If you absorb the pixel data first, you can get insights from the pixel data that human experts can’t. Matt imagines that this idea can be applied to other critical areas like genomics, pathology, etc. He also wants to make this research open source for broader community folks on Azure to benefit from.
“Fairness, reliability, safety, privacy, security… all these things are not afterthoughts.”
Matt shares that Microsoft’s models have learned representations of real health concepts— it understands why it’s being asked XYZ. Acknowledging where a question is coming from goes a long way, so these capabilities are very exciting, but Matt emphasizes that there’s still a big gap between a demo and a real solution that you would put into the healthcare environment. Safety comes first.
“With the two companies working hand in hand very, very close to healthcare, very cognizant of the risk value equation… that’s a force for good.”
He believes we need more of the “boring” stuff in this space: it’s easy to invest in the shiny diagnosis bot, but we need more companies like Guardrails AI that are developing tools from responsible builders.
Microsoft is one of CHAI’s (Coalition for Health AI) lead partners (read Brian Anderson’s episode for more on what CHAI does). Matt shares that he’s excited about the opportunity to have discussions with academic centers, policymakers, and more.
Microsoft tools for startups:
Matt shares that Microsoft has recently launched the Pegasus program to support the pivotal role that startups have in this ecosystem. He strongly believes that giving startups access to tools, like Azure AI tools, are meant to help facilitate and complement the work startups do.
Matt argues that the classic McKinsey model of the Three Horizons may have collapsed a bit. Horizon 3, the creation of new capabilities and new businesses, may actually be closer to Horizon 2, ideas that extend a company’s existing business model to new markets. That may shift your calculation on what gaps startups need to fill.
Matt’s take on startups in the healthcare space
“Startups are the engine of innovation. With this, that is a critical function. Even if they’re wildly successful or not successful at all, we all get amazing learnings from that.”
It’s not just about the technology. Your startup needs to solve a real healthcare problem, meet a real need— and that requires a clinical perspective.
Matt believes that startups often overlook the strength of a multidisciplinary core team. Advisors from the clinical world are great, but they can only go so far. Having someone that’s a champion in the trenches with the team makes a huge difference.
Particularly in the age of AI, Matt believes having the clinician perspective is more important than ever. You’re going to be able to build more rapidly, but you need to make sure to not get off course.
Clinicians are eager to learn more about the AI tech space.
“What’s funny is that I almost feel like I’m at a middle school dance sometimes: there’s the tech folks on one side and the clinicians on the other, they want to dance.”
Matt shares that he has clinicians reaching out to him all the time interested in partnering with startups. They have use cases and are looking for technology partners on the other side. There are different ways to build the relationship— the demand is there.
Matt’s predictions for the future of healthcare AI
Radiology:
Matt shares that radiology has just as big of a burnout problem as any other specialty, particularly because it cuts across all specialties. There are still manual, redundant tasks. The reporting workflow has been optimized since voice-to-text became a thing in the late 90’s, but generative AI has so many more opportunities to accelerate work.
“Can you abstract away some of the things that don’t require a medical degree to do?”
Matt predicts that we’ll see interesting ways to unlock more information from pixel data, get to precision medicine, and take better care of patients.
AI will be used to decrease information asymmetry.
Consider the use cases of taking care of aging parents and kids: people wonder what questions they should be asking their doctor, what medications they should be looking out for. Today, you use inbox messages and wait for your doctor to respond. Matt believes there’s a huge opportunity to decrease information asymmetry as we become more comfortable with having our own health data in our hands. That could look like a companion device or a coach.
Matt also highlights the disparity gaps in healthcare. He wants to see technology used for democratization, bettering health outcomes on a population level.
Fun facts are in the episode only for listening :)
Interested in learning more about Microsoft Health and Life Sciences? Learn more on their healthcare blog, research site, and X.