Innovator Awards Winning Team: Penn Medicine’s PennAInsights
With the twin goals of shifting from episodic to proactive, long-term management of patients with chronic disease and reducing physician burden, radiologists and IT specialists at Penn Medicine in Philadelphia have developed an artificial intelligence solution, PennAInsights, that automates the workflow from image capture through AI analysis to diagnostic reporting. For its potential impact and multidisciplinary nature, Healthcare Innovation recognized this project as one of its three Innovator Award winners for 2025.
As Charles Kahn, M.D., M.S., professor of radiology at the Hospital of the University of Pennsylvania, explains, traditional radiology workflows are labor-intensive and fail to capture subtle early changes. “With the growth in procedure volumes and the smaller growth in the number of radiologists, in many ways, the work outstrips our ability to do it,” he said. “In addition, there is a lot more information in a typical imaging study than any human physician is necessarily going to be able to extract from it. So the notion is that with AI, there is a lot of information that could be extracted from everyday imaging studies that can help support the care of our patients.”
The objective for creating the PennAInsights platform was to develop an automated solution that extracts quantitative measures from existing imaging studies to improve diagnostic accuracy, reduce physician burden, and enable early intervention.
Here is how PennAInsights works: Encrypted images are transferred from the PACS system to a HIPAA-compliant cloud AI server, where a suite of validated AI models process the studies and return quantitative annotations as DICOM structured reports directly into the radiologist’s workflow. In pilot applications, it addresses issues such as abdominal adiposity, liver steatosis, and brain atrophy.
The Penn Medicine radiology and IT execs stress that they were looking to create not just a single application but a platform for other AI developments. The solution integrates seamlessly with existing PACS and dictation systems through a cloud-based, scalable, and secure platform. Its “plug-and-play” modular design allows rapid incorporation of new AI models to address diverse clinical needs.
Walter Witschey, Ph.D., associate professor of radiology at the University of Pennsylvania, explains their thinking. “If you look at the number of FDA-approved AI tools in the field of radiology, it's enormous. But very few of them are actually implemented in clinical care as an end-to-end workflow. They're quite burdensome to use, not interoperable, and not integrated with the health system. None of these solutions really offer what we're looking for, as far as a fully integrated, interoperable AI solution,” he said. “So we thought we would try to build it ourselves. We have the know-how. We’ve been working for a long time with the technical details of images, for instance, the DICOM standard and interoperable connections between devices like the HL7 messaging system. We also had some cool AI applications that we had deployed in research that were very effective. We have deployed them in the Penn Medicine biobank. I think at the time we started, it seemed pretty risky, but as I look back on it now I think we shouldn't have done it any other way.”
One example of a condition this platform has great potential to address is fatty liver disease. “It's important because people who have abnormal fat accumulation in their livers are at increased risk for a variety of conditions. In severe cases, people get tumors in the liver, or they may need a liver transplant,” explained Kahn who also is vice chair of the Department of Radiology.
There is an increasing prevalence of the disease, and people don't necessarily know that they have it, so providing early detection for this condition is an opportunity to catch the disease early and help people mitigate it and potentially save a lot of healthcare costs downstream.
“You'll notice it on a particular study, but not necessarily in its very early form. That's where this AI-based detection can help make us more sensitive and make sure that it becomes a routine part of of the information that's captured,” Kahn added. “So even if you came in for suspected kidney stones, we can look at things like the degree of fat in the liver, or is the liver enlarged? We can look at bone mineral density to see if you might be osteoporotic. We can look for coronary artery calcifications that might indicate that you're at risk for coronary artery disease.”
The project began in 2019 but saw some delays during the pandemic. It went live in May 2023. The effort was multidisciplinary with strong support from IT teams. “From our perspective, it is important to understand the technical requirements and make sure that it goes through the right governance process from an infrastructure and security standpoint and from a data standpoint,” said Anna Schoenbaum, D.N.P., M.S., R.N.-B.C., vice president of applications and digital health at Penn Medicine. “We partner with the providers here to make sure it meets the requirements, to do the testing and to make sure that it is executed safely.”
Ameena Elahi, IS application manager at Penn Medicine, said her team continually seeks innovative ways to eliminate workflow delays by automating processes where possible and maximizing the use of tools within their existing infrastructure.
Pilot study results
In a six-month pilot, Penn AInSights processed 2,973 abdominal CT scans with an average total turnaround of 2.8 minutes. They found that 94.9% of studies were completed in under 5 minutes, ensuring that all AI annotations were available before initial report review. Radiology reports now consistently include quantitative measures that previously appeared in only 24% of brain MRI reports addressing neurodegenerative risks. The pilot workflow operates at approximately $700 per month (less than $1 per patient), demonstrating significant cost efficiency. Early intervention in conditions like fatty liver disease and cognitive decline is expected to reduce downstream complications and overall healthcare costs.
Since the completion of the successful pilot, the platform is now being used broadly across the hospitals that are part of Penn Medicine. The images go to a centralized PACS system and get analyzed by the software. “We have analyzed more than 20,000 studies to date, and we're in the process of expanding to other regions of the body, looking at the chest as well as some tools that will integrate newer advances in AI,” Witschey explained.
As a next step, Penn Medicine clinicians are exploring integration with large language models (LLMs) to automatically structure unstructured radiology report findings—such as detecting adrenal nodules—to trigger clinical decision support in Epic. This extension, currently in a testing environment, promises to further streamline reporting and patient management.
“Using LLMs to structure unstructured radiology report findings will facilitate real-time clinical decision support, streamlines radiologist workflows, enhances diagnostic consistency, and boosts overall efficiency,” Elahi explained.
Kahn said there are even greater possibilities. “We have this amazing resource called the Penn Medicine Biobank. We have the ability to link the various findings that we derive from imaging studies to the genetics and to the other healthcare characteristics — the phenotypes of these patients,” he said. “Here’s where we can connect not just the day-to-day radiology workflow issues, but also connect imaging information to the genetics and gene expression characteristics of our patients, and connect that to their broader health profile. That's a fascinating area.”
Kahn noted that other academic medical centers are also working on applying AI to radiology workflow, but he said Penn Medicine has a lot of the key building blocks in place to make this happen. “You have to have strengths all across the pipeline,” he said. “You have to have a really strong research team — not only Walter and his team, but the Penn Medicine Biobank and the vision that led to that. You have to have a really strong IT team, and we're tremendously thankful for Anna and all the support that she and her team have given us to make this thing possible. And then you have to have people on the clinical side who are interested in bringing this forward to help serve our patients.”
A video overview of Penn AInSights is available here.
