3 Trends Driving Medical AI Democratization
Earlier this summer, I attended the HL7 FHIR DevDays on the Microsoft campus along with over 700 other developers and product managers. I’ve worked with health data standards since the early ’90s, and this event was exceptional in its focus on implementation. There was even a breakout session explaining the alignment of standards bodies around FHIR and DICOMweb. A few weeks later, at the annual Society for Imaging Informatics (SIIM), it was a pleasant surprise to learn the latest HL7 FHIR and DICOMweb standards are required to access the SIIM-ACR Pneumothorax competition data. These steps are nudging the model developer community into using standard data access methods, helping to lay a foundation for greater interoperability. The move toward interoperability standards has been glacial in the healthcare industry, but new regulatory requirements and increasing uptake by technology vendors make this a positive signal.
Open-source high-value use cases have been made available by the American College of Radiology (ACR) to help guide developers toward clinically relevant AI. The role of data science competitions has long shaped the industry. This summer’s Pneumothorax Kaggle competition sponsored by SIIM and ACR illustrates the power of combining well-annotated data sets with teams eager to win glory (and $30k). The ten winning competition entries will become available to the world via open-source licensing (MIT, BSD, or Apache) for any purpose, including commercial use. This has several exciting—and for some AI vendors, disruptive—implications. In past Kaggle competitions, some winning teams formed companies to commercialize their intellectual property. The winners of this competition may do the same, but their models are not exclusively theirs. What does this do to the nascent marketplace for medical imaging analytics models?
There is much more to creating a software medical device than just the algorithm. Innovative business models will be required by medical AI development companies and their clinical partners when high-quality open-source models become a major trend.
ACR introduced its new ACR AI-LAB as a “data science toolkit designed to democratize AI.” ACR has implemented a cloud-based set of tools that can be used directly by radiologists to “Do It Yourself (DIY)” and develop their own AI-powered solutions within their institutions. The ability of radiologists and other specialists to get hands-on experience with AI development tools can demystify AI and create evidence of clinical value for patients. Other free research tools, such as NVIDIA CLARA, also illustrate this trend. This builds on earlier contributions by technology giants who have open-sourced many of the tools of the trade. It will be informative to see how rapidly such annotation and algorithm testing tools are adopted. We are especially interested in following clinical research efforts by medical centers and large radiology practices which are now engaging in this journey.
The popularization of AI development spurred by these trends will have both intended and unintended consequences. Some of the intended consequences are already visible, with 1,287 teams competing in the Pneumothorax competition. Similar competitions are in the wings. This trend may encourage healthcare delivery organizations to develop models that are locally optimized to their patient populations. Unintended consequences may be the proliferation of home-grown AI models that are implemented under the FDA’s radar. Like Lab Developed Tests (LDTs) in widespread use today at medical centers, it is easy to imagine rapid deployment of home-grown AI algorithms without FDA review. This new era requires new methodologies and supporting quality systems to keep up with the pace of innovation.
CuraCloud is introducing new customized services to support healthcare organizations and vendors in developing and validating their own AI software medical devices. We enjoy helping others create the next generation of medical AI.