How democratizing domain-specific published AI models on top of cloud infrastructure could create a performance-based business model for AI in healthcare
Artificial Intelligence (AI) in healthcare has become the subject of both great promise and great hyperbole. Beyond buzzwords and a plethora of VC investments, AI and other mathematical techniques are beginning to emerge in a second wave of domain-specific “systems of intelligence.” The key missing factor has been a business model in the payer and provider community that enables the “best” (aka: most validated, most clinically proven, most workflow integrated) models to drive an economic value both to the care paradigm and to the cost centers of healthcare data storage and analytical processing.
A Bit of Background on Healthcare vs. The World of Big Data
I’m often asked by colleagues from other industries (retail, finance, industrial, advertising, etc) why healthcare has been such a laggard in the use of data for AI-driven insight and decision making. The easy answer is to point at regulation, HIPAA, GINA, the FDA, and even more importantly the role of expertise as driven by the training and judgement of well trained physicians as factors that have slowed AI adoption. Data exists in protected, separated silos with huge difficulties around unstructured information and complex disease contexts that encompass high stakes choices for nuanced differences between patients and outcomes of treatment. The hard answer, that those of us who have lived healthcare from the business side know, is that a business model that rewards highly accurate recommendations and driving insights to physicians and patients just doesn’t exist. If a grad student with a passion for predicting COPD from lung volume measurements and EKG readings (Wadhwani et al.), or a corporation with a predictive model for identifying pre-tumor formations from imaging data has no real place to take these predictions, then they sit mostly in interesting academic studies with little true impact on the care of patients, particularly outside of the ivory tower research institutions.
The majority of AI technique-driven predictions have gone through the normal channels of peer review, companion diagnostics, or software as a device regulatory review. While medical scrutiny is critical, and there is actually real danger in deploying “black box” predictions, it does seem to miss the performance-based meritocracy that governs the use of AI in most other industries. In advertising for instance, real-time measures like “click rate” and “conversion” drive what is almost a marketplace for bid-ask on algorithm performance. In finance, algorithms are judged on daily performance, and derivative optionality can make or break fund performance where the scoreboard is kept daily and performance matters. There is no such simple, measurable calculation in healthcare today.
Cloud Vendors are Beginning to Create a Model for Performance-based AI in Healthcare
As the cloud storage wars have entered healthcare, a question has emerged: beyond security and the improved economics of cloud storage vs. physical storage what will drive revenue for cloud vendors? Further than that, how will horizontal cloud vendors prove validity of AI offerings to impact business needs of providers and payers such as improved quality, decreased workload, and better outcomes? From IBM Watson to Microsoft, Amazon, and Google, the question of how AI practically impacts care is critical for most healthcare players. However, there seems to be a model emerging that could advance domain-driven AI though these platforms into practice based on performance and, perhaps, paid by usage.
The Cash Register: AI.ML Compute Cost. How Microsoft and Others Could Democratize AI in Healthcare
Artists who submit content to Spotify or iTunes are inherently compensated based on the listens to their music. If the audience deems it quality, then compensation flows back to the creator of the content. Netflix, YouTube, and others have established this same model. What is fascinating is that as healthcare players have begun to transfer their data cloud storage, a major part of their suite of services includes AI platforms like Azure, that charge based on GPU compute that is run when AI models process data. While it is early, this creates the potential that as more AI models are developed and run with greater frequency, perhaps even on every patient, compute will represent the unit on which this value could be measured. Call it “Bitcoin for Healthcare Insight,” or maybe more accurately “User Votes on Which AI Model is Most Valuable.”
For AI model developers in healthcare this could mean that, whether you are a grad student or a multinational company, your model could have a marketplace by which you’ll succeed if your accuracy and utility actually drive a financial component of the data economy in healthcare. Further, this puts healthcare providers and payers in the position of “voting” for AI models that actually positively impact their business and patient outcomes. Imagine an oncologist with “playlist” of AI models they might utilize for a subset of patients, trained and learning on new data and transparently judged on performance. This type of model puts the physician and administrator in a critical role of judging utility based on output and performance to their patients and their business.
Microsoft and Others in an Exciting Evolution as an AI marketplace
Who is moving most quickly to create this new kind of AI model marketplace? Microsoft just leapfrogged Alphabet/Google last week to become the third most valuable company in the world – only behind Apple and Amazon. How have they achieved this? Dave Gershgornfrom Quartz reported:
“Microsoft is doing great. At a spry 43-years-old, the company has picked up momentum in the market. Microsoft’s market capitalization gained about 40% over the last year, buoyed by its successful software subscriptions business and cloud computing growth.
It’s a period of bold expansion for the technology company and also an opportunity for healthcare innovators who can partner with Microsoft in their focus on enterprise growth.
Our parent company, Symphony AI, just announced a fund-wide partnership with Microsoft to develop next generation AI applications and domain-specific “systems of intelligence” built on the Microsoft application stack and making them available to the universe of Microsoft Azure-based users. While there’s business and strategic reasons to align ourselves to Microsoft’s continued innovation, both in healthcare and artificial intelligence (AI) technology, the real potential is the ability to deliver meaningful domain-specific AI driven insights to a wide variety of healthcare users.
What is perhaps most exciting to me is that, despite years of siloed data preventing healthcare experts from answering even basic questions, Microsoft’s movement signals that we have now entered the stage of AI-driven insight. AI is no longer relegated to academic literature, or complex custom deployments. Through teaming with Microsoft, we will be able to make this work mainstream:
These examples of our work to use AI to advance oncology care are now API- and AKS-ready and available within the Microsoft technology stack. They provide transparent patient level recommendations and predictions about the outcomes of patient care and are designed to fit within the workflow of existing healthcare and life science companies.
AI technology is ready to prove itself in healthcare. There is have a decade of testing, experimentation and advances. New studies on AI in clinical use, such as the recent ASCO research showing our pre-trained modules can successfully predict chemotherapy-induced neutropenia and make cancer treatment safer, are emerging everyday. With Microsoft and the future emergence of a Spotify-style model marketplace, it is possible for AI to finally become mainstream. To help physicians and patients in the everyday, life-and-death decision making that moves healthcare forward.