Did you miss a session at the Data Summit? Watch On-Demand Here.
Follow along with VentureBeat’s coverage from Nvidia’s GTC 2022 event >>
Digital twins play a crucial role in Nvidia platforms for developing robots and self-driving cars. But medical digital twins use cases face several significant regulatory, technological and privacy hurdles. At the GTC this year, Nvidia is showcasing a variety of significant advances that could drive the adoption of digital twins in medicine.
The hardware supporting AI has matured to the stage where the AI industry needs to be continuously improved, updated, tested, and validated. “We will not be able to scale that without digital twins,” Nvidia’s VP of healthcare, Kimberly Powell, said in a press conference.
Key healthcare advances introduced at GTC include synthetic data generation, the commercial release of its Clara medical AI platform, an enhanced DNA sequencing workflow, improved pharmacovigilance capabilities and improved drug discovery tools. Enhanced digital twins capabilities will eventually take advantage of these advances to dramatically improve patient safety and support new business models in the healthcare industry.
Enhancing robotic data platforms
Nvidia has been an innovative leader when it comes to new AI platforms for autonomous driving (Drive), robotic design (Isaac) and healthcare (Clara). These platforms simplify the development, deployment and continuous improvement of AI for industry domains. Each platform includes all the capabilities required to bring data from disparate sources into a development environment, train new algorithms at scale and then deploy them to new inferencing hardware.
Healthcare data accounts for 30% of global data requirements and is growing at 36% CAGR. Nvidia Clara streamlines AI workflows with AI training; over 40 pre-trained models, applications and platforms across data centers; in standalone servers; or integrated into medical devices. However, Nvidia has not announced a specific medical digital twin capability yet.
In contrast, both Drive and Isaac include digital twins capabilities that simplify workflows around product design, AI training workflows and testing the performance of new combinations of hardware and software. For example, Drive supports complete simulation environments that can synthetically generate variations that reflect the impact of rain, snow and darkness. These simulations can help train and develop models and predict when a model may not perform well enough. Similarly, Isaac helps build and test new robotic hardware and algorithms in these simulated environments.
Digital twin capabilities are more challenging in medicine due to privacy safeguards, medical regulations and safety considerations. Although companies address these concerns in one-off implementations today, these are difficult to scale. The combination of Nvidia’s existing toolchain and the recent announcement could help address these challenges.
Healthcare announcements at GTC
- Ultrarapid nanopore analysis pipeline (UNAP) is a new DNA sequencing platform that runs on a single DGX A100 to reduce the compute cost for sequencing a whole genome from $568 to $183. It recently helped set the world record for sequencing an entire genome in four hours and ten minutes.
- Support for four startups developing AI transformers for decision-making, therapeutics and drug discovery using the fastest supercomputers in the UK. Transformers help teams analyze unlabeled data, representing the bulk of medical data.
- Early access to Megamolbart, a new training framework and generative model for chemistry developed in partnership with Astrazeneca. a natural language processing (NLP) model that reads the text format of chemical compounds and uses AI to generate new molecules. The transformer chemistry model can train chemical language models with over 1 billion parameters using the Nvidia Nemo Megatron framework.
- A new domain-specific NLP model from Janssen built on Biomegatron improves adverse pharmaceutical event detection by 12%, for a total detection rate of 88%.
- SynGatorTron, the world’s largest clinical language model generation tool, in partnership with the University of Florida. It automates synthetic data creation of healthcare data to improve AI models while protecting privacy. It could be used to create digital twins for patient records as a control group in clinical trials. A complementary GatorTron model can also enhance medical chatbots, biomedical research, clinical trial matching and medical event detection.
- Nvidia Clara Holoscan MGC, a reference design for a real-time medical-grade AI computing platform scheduled for early access in Q1 2023. It promises to industrialize AI development and comes with a 10-year software stack support agreement. Early hardware partners include ADLINK, Advantech, Dedicated Computing, Kontron, Leadtek, MBX Systems, Onyx Healthcare, Portwell, Prodrive Technologies, RYOYO Electro and Yuan High-Tech.
Powering new business models
Powell expects that current breakthroughs in synthetic data generation and improved modeling tools will also help bring digital twins to healthcare. For example, past work done with the Cambridge One supercomputer helped the King’s College London generate synthetic brains. This allowed them to test new algorithms and represent populations in different ways.
Healthcare could also adopt business models similar to those that innovators like Tesla have brought to the automobile industry with SaaS offerings for autonomous driving. In the traditional model, car companies would sell a car and make all their money from the direct sale of the vehicle. Tesla came along with a new programmable AI computing platform that continuously updates new features and improves over time.
Similarly, healthcare companies could develop innovative new enhancements to existing tools like MRI scanners or endoscopy robots that improve medical procedures and clinical workflows. Powell said, “We need these capabilities to be married with the new devices. The additional cost to build these is tiny compared to the economics of the services you can build on top of it.”
Powell predicts that the current work on weaving digital twins into healthcare AI workflows is only the beginning of at least a ten-year journey. And beyond that, it will become a critical part of AI and product development in healthcare.
“It will be synonymous with how AI is so pervasive in everything we do today. Digital twins will be pervasive in everything we do across every industry across the next decade,” Powell said.