Q&A: NVIDIA Healthcare Leader on What It Takes To Make Edge-to-Cloud AI Work

HEALTHTECH: What kind of technology and governance foundation is needed to deploy edge-to-cloud AI? Do most healthcare organizations already have this foundation in place?

NIEWOLNY: New software applications are driving this need in many cases. Some healthcare organizations have bits and pieces of that infrastructure and governance today, but you’re looking at a foundation of accelerated compute. You need data center and edge in addition to size for the workload. You need reliable networking, storage, standardized containers, application programming interfaces, identity access, device management and monitoring. There are a lot of pieces that a standard healthcare IT system should have in place.

However, when you start bringing in mission-critical applications, you now need a governance system that connects the clinical with IT, security, privacy, legal and the development teams that may be building and deploying AI. 

Governance models are being developed, and some of this is being outlined by regulatory agencies. This is true especially in terms of what they’re expecting from device manufacturers and software providers, such as predetermined control-change plans and ensuring they can maintain and update the software on a regular basis. 

HEALTHTECH: How can hospitals balance the need to keep data local with the benefits of a centralized AI model?

NIEWOLNY: That’s tricky, and I’d say they should definitely err on the side of caution. Most people don’t want raw patient data to leave the site. In that case, you want models that can be deployed locally at the edge. Then the organizations end up choosing only to centralize what they deem is appropriate, such as approved model artifacts, de-identified data, aggregated telemetry data or performance metrics. That’s the type of data that can be managed off-prem in a cloud or data center.

In many cases, we’re working with organizations to minimize some of their data. For some of the larger healthcare organizations we’re looking at building federated data approaches and creating synthetic data on-prem to make sure they have that data in place. Centralization comes into play when you start talking about large-scale governance, reproducibility, fleet health, learning across multiple sites and anything involving a lot of patient data within the confines of a healthcare organization.

RELATED: Enable faster clinical decisions at the point of care with real-time analytics at the edge.

HEALTHTECH: What advice do you have for organizations interested in deploying edge-to-cloud AI?

NIEWOLNY: Health systems looking at edge-to-cloud deployments are often buyers of technology, not necessarily builders, so they’re going to have to understand the different AI applications they plan on running. Some med tech and health tech may provide their own infrastructure as part of an end-to-end solution due to the complexity of working within a health organization’s IT system. However, some of them may require the organization to have their own edge infrastructure in place.

Ask how edge compute is being delivered for the AI applications and software you’re leveraging. Is that being delivered with the solution, similar to a medical device, or is this something that is truly being deployed as software? We are working with different healthcare organizations to help them build a true infrastructure from the ground up that allows them to provide edge-to-cloud capabilities within the hospital — essentially, at the time of build or retrofit.

NVIDIA is working with a lot of these companies to democratize edge technology, which can reduce costs. Providing edge architecture based on the software you’re using and the workloads you need, you can ensure your hospital is future-proofed with edge-to-cloud applications that are true software applications running on your IT infrastructure. 

HEALTHTECH: How can healthcare organizations ensure the security of these tools? What do they need to know about managing and monitoring them?

NIEWOLNY: Security needs to cover the full life cycle of an edge device. These devices need hardware root of trust, secure boot, signed software, model artifacts, network segmentation and encryption, all embedded deep within the hardware and software. These are the components that are as close to the patient as possible and should be treated just as you would treat a medical device.

The organization needs to complete a full inventory of devices and applications while monitoring all the models, versions and data flow. It needs to be designed from the ground up. This should be managed as part of the IT network. Organizations need to continuously observe latency, resource health, input quality, output distribution, model drift and, most important, outcomes.

HEALTHTECH: How does NVIDIA help healthcare organizations with their edge-to-cloud AI initiatives?

NIEWOLNY: NVIDIA is instrumental in working with med tech and health tech companies on their AI model development. From that, we have taken a corporatewide stance to index heavily on accelerating AI model development using open models. 

We’ve introduced initiatives related to robotic surgery and spatial intelligence to address some of healthcare’s key challenges. As part of the Open H initiative, we’ve pulled together a number of academic and med tech partners to assemble the world’s largest data set of robotic surgical data. We’re also doing the same thing for some spatial intelligence applications, such as patient and operating room monitoring.

These initiatives give us the ability to move to another area that we’re heavily invested in: simulation. We’re creating synthetic data sets to augment these data sets, creating good foundation models for use cases that our customers and the industry can fine-tune and build on top of. It helps to accelerate their ability to build models, which companies then can turn into applications. 

We provide all the medical-grade hardware needed to deploy at the edge in healthcare. We work with a lot of the major original equipment and design manufacturers. You may be familiar with some of them, such as the Jetson platform, IGX platform and DGX Spark, which is widely used at the edge in a lot of medical devices. We accelerate model development and testing with our simulation platforms. 

Finally, all of this ends up getting deployed at the edge, on a unified and common architecture that is provided by NVIDIA yet sold to the market through a number of our OEM and ODM partners. 

HEALTHTECH: Is there anything else you want to mention?

NIEWOLNY: Something that people often want to know is, why is the time now? We’ve been talking about AI for quite a few years, but now forces are coming into a perfect convergence. One is that the industry is running into some very serious capacity problems related to staffing shortages. We’re also always talking about the cost of care. When you’re looking at ways to improve access to care while dealing with a capacity problem, you’re going to look to tools such as automation and AI to help solve those problems.

The second piece to this, where NVIDIA plays a key role, is that foundationally, all these med tech and health tech companies have had to build full-stack computing architectures for each of those areas that I mentioned: AI model training, simulation and deployment. Now you’re starting to see more collaboration between both the technology companies and healthcare organizations. They are collaborating to accelerate the development of this technology, which leads to faster deployment. 

The challenge has become painful enough that it has forced the industry to collaborate at the exact same time that these tools are being developed. To me, it’s an exciting time. I think we’re going to see a once-in-a-generation platform shift that is truly being pulled by healthcare, not pushed by tech companies.