After retiring from being the first President of Oracle’s cloud computing business, I returned to Stanford University to start the first class on cloud computing, CS309A. Several years ago, a unique student enrolled. That student — who after graduation went on to be a leader in AI in medicine, was Dr. Anthony Chang. It was through him that I was more than a little surprised to learn, even with all that technology had to offer, the health care system was (and still is) using CD-ROMs (or the “more advanced” USB sticks) to transfer health data. Furthermore, I learned the rapid advances and accuracy we’ve seen in consumer AI was simply not happening in medicine, in large part because improvements in accuracy are inherently dependent on access to large quantities of diverse data.

During COVID, I made the decision to come out of retirement and assemble a team of experts. Our mission? Create privacy (and life) preserving, real-time applications based on access to data generated in all 1,000,000 healthcare machines in all 500 children’s hospitals. Success in this mission will reduce healthcare inequity, lower healthcare costs, and improve patient outcomes not just nationally, but globally. We call this the Pediatric Moonshot.

While a centralized architecture approach — involving the gathering of vast amounts of data into centralized repositories — has been used quite successfully to speed the development of consumer AI applications such as ChatGPT, this approach will not work for AI applications in medicine. Centralized architectures are not the answer for AI in medicine for several reasons. Most notably, centralized architectures are not network preserving, not application friendly, not real-time, not privacy-preserving, and not sovereignty preserving. So, like the original moon shot, we needed to develop a new rocket.

The Bevelcloud team designed a decentralized architecture for AI in medicine. We call this decentralized, in-the-building infrastructure an edge cloud. It must operate in the building as the healthcare machines are in the building, which means security management and privacy management were built right into its core. Furthermore, the edge cloud must operate on all six continents, where each of the 500 children’s hospitals are located. Since the mission is to provide access to data in all 1,000,000 healthcare machines, we have created healthcare machine digital twins to standardize access to the data from imaging machines, blood analyzers, and bedside monitors in real-time. The architecture also supports PACS digital twins to standardize access to offline image data as well as EMR digital twins to standardize access to medical records.

Using these common data standards application developers can use BevelCloud Studio to build privacy-preserving, real-time applications which can run anywhere in the world. The closest analogy to this approach is what Apple did with the iPhone. By creating a platform, standardizing access to the camera and GPS, they encouraged thousands of independent developers to build applications. When the iPhone launched, who could have imagined the likes of TikTok, Instagram, or WeChat? Similarly, we imagine thousands of creative developers building a wide variety of applications in cardiology, radiology, oncology, orthopedics, emergency medicine, rare diseases, neonatology, and neurology.

There are many applications possible. One, which we have focused on, is an emergency medicine image-sharing application. From a recent paper on the state of the art in radiologic image sharing among U.S. children’s hospitals : “The prompt and secure transfer of imaging is vital for patient safety as demand for imaging increases.” Many pediatric emergency room physicians validate this statement, identifying image sharing as a universal problem, especially when dealing with radiologic images obtained from sending hospitals or clinics. At times, the only hope of accessing these images depends on whether or not the family took a picture of the screen.

It does not have to be this way. Building a nationwide or global network to share scarce pediatric clinical expertise with clinics and non-children’s hospitals is within our grasp. In 2022, our team demonstrated real-time global image sharing. Dr. Chang brought his eldest daughter, who has a congenital heart problem, into the echocardiology lab at Children’s Hospital in Orange County, California. Her ultrasound was shared — in real-time, as it was being conducted in California — with experts at Bambino Gesu in Vatican City. Check out a short video from that historic day.

Recognizing that pediatric expertise is far too scarce for those children who are not geographically or socially lucky enough to have direct access to a pediatrician or other child health professional also justifies the need to build AI for children’s medicine. Again — providing high-quality, accurate care can be better achieved, despite geographic and socioeconomic considerations, through deploying privacy-preserving, real-time, AI applications to the point of care at a children’s hospital, non-children’s hospital, or clinic. Deploying cardiology, neurology, or orthopedic applications to the point of care is simple using the edge cloud platform.

What’s hard is how to train accurate, non-biased AI applications while preserving privacy. Luckily, a technology called federated learning has emerged to train consumer AI applications, while still preserving both network resources and privacy. Siri is a great example of the use of federated learning in consumer AI. Siri can learn every time you talk not by sending your voice data to a centralized service, but rather by learning from your voice data on and across a highly decentralized architecture — millions of iPhones.

Training accurate, non-biased AI applications in medicine requires access to large quantities of diverse data, and this data exists. For example, consider that the echocardiology labs found in all 500 children’s hospitals around the world cumulatively produce over 6,000,000 TB of data each year. That’s 150,000 times the size of the NIH centralized Imaging Data Commons. With the in-the-building edge cloud, we can employ federated learning to train highly accurate AI applications on every possible pediatric cardiology condition while preserving privacy.

Since developing AI applications without data is like building enterprise applications without a programming language, our objective is to build a Federated Learning Lab for Children’s Medicine. Initially, the Lab will be focused specifically on pediatric cardiology, beginning with 100 edge servers “twinned” to 100 ultrasound machines in six locations on two continents. By providing an open-source AI application to measure ejection fraction governed under a multi-site, multi-country IRB, we will be able to experiment with various software stacks and federated learning strategies. This is key to the next step in AI in medicine.

It’s been estimated it took 40,000 people to reach the moon in the original moon shot. The Pediatric Moonshot will also take a community. Success in the mission will reduce healthcare inequity, lower healthcare costs, and improve patient outcomes nationally and globally. Fueling our efforts is our firm belief that software can — and should — be used to do good in the world.

We believe in changing healthcare one connection at a time. If you are interested in the opinions in this piece, in connecting with the author, or the opportunity to submit an article, let us know. We love to help bring people together! [email protected]