Real-World AI in Healthcare: Inside the Cutting-Edge Healthcare at Penn and Sutter Health

Updated: May 03 2025 16:45

AI Summary: Leading healthcare institutions like Penn Medicine and Sutter Health are implementing groundbreaking AI initiatives, moving theoretical applications into real-world solutions to transform patient care. Penn Medicine leverages AI in areas like radiology for faster disease identification, drug repurposing for finding new uses for existing medications, and the Observer Project to revolutionize clinical documentation and reduce physician burnout, facilitated by the Penn Center for Innovation and a collaborative ecosystem.

The healthcare industry stands at the precipice of a technological revolution, with artificial intelligence poised to fundamentally transform how we prevent, diagnose, and treat diseases. In parallel with the 2025 JP Morgan Healthcare Conference, Penn Center for Innovation (PCI) hosted its annual Innovation@Penn program in San Francisco. The program brought together professionals from across the pharma, MedTech and biotech industries, and investor community, and provided a platform for in-depth discussions about the latest cutting-edge technologies and their practical applications. Below is the full video if you are interested:


There are also some interesting AI initiatives at Sutter Health, where researchers and clinicians are turning theoretical AI applications into real-world healthcare solutions that benefit both patients and providers. Let's dive into how these pioneering organizations are integrating AI into healthcare and what it means for the future of medicine.


At the heart of Penn Medicine's success in AI implementation is the Penn Center for Innovation (PCI), which serves as a crucial bridge between academic research and commercial application. Acting as a "one-stop shop" for industry engagement, PCI has played a pivotal role in advancing Penn's national ranking in annual licensing income and executing numerous commercialization agreements over the past decade.

What makes Penn's approach distinctive is their strategic focus on impact and innovation, coupled with a genuine appreciation for commercial sector engagement. This vision has positioned Penn uniquely to make significant strides in applying AI to healthcare challenges, leveraging their exceptional leadership and interdisciplinary opportunities.


Of all medical specialties, radiology has perhaps seen the most dramatic transformation through AI implementation. According to Walter Witschey, Associate Professor of Radiology at Penn Medicine, a substantial percentage of FDA applications for AI in medicine are concentrated in radiology.

"AI is fundamentally changing clinical care by helping to identify important indications with greater precision and earlier in the patient journey," explains Witschey. The integration has been remarkably seamless, with tools deployed to help radiologists interact with AI naturally during their workflow.

A standout success story is Penn's implementation of an AI tool that analyzes CT images to identify conditions like hepatic steatosis (fatty liver disease). This tool has already screened over 15,000 patients, identifying conditions more quickly and effectively than traditional methods.

Perhaps most exciting is the emergence of vision language models, which promise to transform radiology by generating explainable information from medical images that can be understood by both specialists and non-physicians alike. These models represent a significant leap forward from the convolutional neural networks that began transforming radiology about five years ago.


Another area where Penn Medicine is making remarkable strides is in drug repurposing using AI algorithms. David Fajgenbaum, a physician-scientist at Penn with a personal connection to this approach (having benefited from drug repurposing himself), has pioneered what he calls "computational phenomics" – using AI to systematically identify promising repurposing opportunities across all drugs and diseases.

The impact of this approach can be life-changing and immediate. Fajgenbaum shared an example where a repurposed drug, identified through AI prediction, saved the life of a patient with a rare disease within days of the algorithm's suggestion.

"With thousands of FDA-approved drugs and a vast number of diseases without approved therapies, AI offers a powerful way to find new uses for existing molecules," Fajgenbaum explains. The most exciting prospect is the ability to analyze all possible drug-disease pairs to identify interventions that could save lives globally.

What's particularly impressive is the acceleration of this process. What previously took 100 days to compute just two years ago now takes only 17 hours, allowing for continuous iteration and refinement of the models.


Perhaps one of the most transformative applications of AI at Penn Medicine is "The Observer Project," led by Dr. Kevin Johnson and highlighted by Marylyn Ritchie, Vice Dean for AI and Computing at the Perelman School of Medicine.

This innovative project involves recording primary care appointments (with patient consent) and using AI to:

• Extract structured information for electronic health records
• Generate clinical notes from both patient and provider perspectives
• Analyze patient sentiment, gait, and posture

The potential impact on physician burnout is substantial. By automating documentation, doctors can focus more on patient interaction rather than paperwork. Additionally, the project promises more accurate and consistent data for research, potentially transforming precision medicine.

"This represents an evolution from ambient listening to ambient sensing," explains Ritchie, "capturing more dimensions of the clinical interaction than ever before."


A recurring theme in Penn's AI initiatives is the breathtaking pace of advancement. Walter Witschey notes the progression from rule-based vision systems to convolutional neural networks, and now to sophisticated vision language models in radiology – all within a span of just five years.

In drug repurposing, computation times have dropped from months to hours. And in research assistance, the capabilities of large language models like ChatGPT have improved dramatically in just the past year, moving from providing often inaccurate information to generating highly accurate content that significantly accelerates the research process.

This accelerating pace creates both opportunities and challenges, requiring adaptable governance frameworks and continuous reassessment of implementation strategies.


Despite the excitement surrounding AI in healthcare, researchers and clinicians at both institutions remain clear-eyed about the challenges involved, particularly regarding trust, safety, and fairness.

Marylyn Ritchie at Penn Medicine shared a cautionary tale about an AI system for melanoma detection that failed in two critical ways: it learned to identify rulers in images rather than melanomas themselves, and it performed poorly on different skin tones due to biased training data.

"AI is fundamentally pattern-recognition technology," Ritchie explains. "It will learn whatever patterns are present in the data, intended or not." This underscores the importance of diverse training data that represents the full spectrum of populations being served.

Walter Witschey highlighted another example in radiology, where algorithms for detecting conditions like pneumothorax showed reduced performance in certain demographic groups, such as women and older individuals. "We need greater appreciation for including data from diverse populations across different geographical areas to build more robust and equitable models," he emphasizes.

At Sutter Health, patient trust is equally paramount. Kiran Mysore acknowledged that some patients might be hesitant about AI in clinical settings: "Even for me, when I see my doctor, I know that he's using technology. I start to wonder, is this tool capturing the right things? So that gets into a bigger conversation about trust." To address this, Sutter ensures patients are informed and asked for consent before AI tools are used.

Guru Sundar from Abridge explains their approach to building trust: "It's not perfect and we don't claim to be perfect, but we do strive for it. That's where the feedback mechanism comes in. As physicians use our platform, we capture the edits they make to AI-generated notes before they're finalized in the medical record. We use that data to improve our model."

Across both organizations, researchers and clinicians universally stress the importance of keeping humans in the loop to evaluate AI predictions and ensure they make sense in clinical contexts. As Mysore puts it:

AI is successful when there's mutual trust between a patient, a physician and the predictions. That trust is built over time.

What makes Penn Medicine's AI initiatives particularly successful is the unique ecosystem they've cultivated. Several key factors contribute to this environment:

• The seamless connection between algorithm development, laboratory validation, and first-in-human clinical trials within a tightly linked School of Medicine and Health System
• Co-location of various schools (medicine, engineering, communication, business) on a single campus, facilitating interdisciplinary collaboration
• Initiatives like the Colton Center for Autoimmunity that specifically bring together leaders in AI, wet lab research, and clinical research
• Leadership that includes many individuals with both MD and PhD degrees focused on translating basic science into patient benefit
• A culture that embraces calculated risks and positions leaders as generators, not just receivers, of innovation
• Resources like the Penn Medicine Biobank, providing rich data (EHR records, biospecimens) for developing and testing AI models

David Fajgenbaum particularly emphasized Penn's exceptional strength in translational research – "the ability to translate ideas into actual impact for patients" – attributing it partly to the dual-trained leadership that understands both clinical needs and technical possibilities.


Looking ahead, Penn Medicine's researchers identified several exciting frontiers and opportunities for collaboration:

1. Advancing promising drug repurposing discoveries that lack commercial incentive, such as using leucovorin for a rare subtype of autism or lidocaine to prevent breast cancer recurrence
2. Deploying promising algorithms in robust patient populations using the Penn Medicine Biobank's extensive dataset
3. Scaling deployed AI tools in radiology beyond specific procedures to encompass all imaging studies, which requires significant engineering and cloud computing expertise
4. Reimplementing tools developed at Penn in other hospitals to ensure they work reliably in diverse settings
5. Developing safe and ethical ways to share and de-identify data while protecting patient privacy, potentially through federated learning approaches

The researchers emphasized that external collaboration can help accelerate the pace of research beyond the often slower timelines of traditional grant funding, creating opportunities for startups, researchers, and investors to engage with Penn's ecosystem.


While Penn Medicine is breaking new ground in research applications of AI, Sutter Health in Northern California is demonstrating how thoughtfully implemented AI can enhance the human connection in everyday clinical practice.

"What do AI and air conditioning have in common? When they're doing their jobs right, you don't even notice — but the difference is definitely there," says Guru Sundar, Vice President of Marketing at Abridge, an AI-powered medical scribe company partnering with Sutter Health. This analogy perfectly captures Sutter's philosophy toward AI implementation: technology that works quietly in the background while amplifying the human aspects of healthcare.


One of Sutter's most successful AI implementations has been the deployment of AI-powered scribes that document patient visits through the Abridge platform. Dr. Alice Woo, a plastic and reconstructive surgeon at Sutter West Bay Medical Group who sees 45-50 patients daily, puts it bluntly: "It's given me my life back." The results have been remarkable:

• 78% of Sutter clinicians reported "significant improvement in their job satisfaction" when using AI scribes
• 49% reported a reduced "cognitive load"
• Nearly 60% felt the quality of their clinical notes improved

"It elevates the patient experience," explains Kiran Mysore, Chief Data Analytics Officer at Sutter Health. "I have a stronger relationship with my physician because this person is actually interacting and responding to what I'm saying instead of typing out notes."


Beyond clinical documentation, Sutter Health has pioneered using AI to improve access to critical preventive care screenings, particularly for vulnerable populations.

A standout example is their innovative approach to diabetic retinopathy screening. This sight-threatening complication of diabetes requires annual eye exams, but historically, nearly half of patients who needed one didn't get it due to accessibility barriers. Traditional exams required eye dilation (preventing driving afterward), visits to specialists often concentrated in urban areas, and navigating appointment availability and insurance coverage.


Sutter's solution? AI-enhanced retinal cameras placed directly in primary care offices that can:

• Capture non-dilated retinal images during routine primary care visits
• Use AI to detect even early signs of retinopathy in under 60 seconds
• Eliminate the need for separate specialist appointments for screening

The impact has been dramatic. During a six-site pilot program:

• 92% of completed diabetic eye exams were performed in primary care using the AI-enhanced cameras
• Some practices saw a 12-fold increase in completed diabetic eye examinations
• Overall compliance with this critical screening jumped from 62% to 68% system-wide

"This project not only demonstrates Sutter's commitment to expanding access, meeting patients where they are, and removing friction from the healthcare experience; it shows the positive result that these changes can have," says Dr. Kristen M.J. Azar, Executive Director of Sutter's Institute for Advancing Health Equity.


Sutter Health is also leveraging AI to improve cancer detection. Many Sutter gastroenterologists now use "GI Genius," an AI-powered endoscopy module that assists in identifying colorectal polyps during colonoscopies.

"It's a tool that strengthens our ability to find potentially cancerous polyps early," explains Dr. Ashwini Anumandla, who has used the technology at Sutter's Stockton Surgery Center for about two years. "It doesn't replace the human connection in medicine, it just improves our efficiency and effectiveness."


The technology functions as a "second pair of eyes," scanning the tissue lining of the colon during procedures and highlighting suspicious areas with a digital green box on the screen. Studies show that AI-assisted colonoscopies enhance polyp and adenoma detection rates, with every 1% increase in adenoma detection translating to a 3% decrease in colorectal cancer risk.

Dr. Kuntal Thaker, who has performed nearly 2,000 colonoscopies using GI Genius, noted: "The technology helps us spot lesions that otherwise may have been missed. In medicine, we always want to be improving. This is a tool that makes patient care better."


As demonstrated by both Penn Medicine and Sutter Health, AI is already moving healthcare upstream toward preventative care, though opinions vary on whether its current implementation is overhyped or underhyped – indicating areas ripe for further exploration and debate.

What remains clear is both institutions' commitment to facilitating partnerships that advance AI in healthcare. The collaborative spirit and dedication to translating discoveries into real-world impact define these healthcare ecosystems and signal a promising future for AI in medicine.

As we stand at this inflection point in healthcare technology, the work being done at institutions like Penn Medicine and Sutter Health demonstrates that AI is not just a theoretical tool for the future – it's already transforming patient care in tangible, measurable ways. The challenge now is scaling these innovations responsibly, ensuring they benefit all patients equitably, and creating governance frameworks that can keep pace with the rapid evolution of the technology.