By Mike Miliard, HealthcareIT News | November 29, 2018
The most advanced machine learning algorithms won’t get health systems where they want to go unless the data fueling them is high-test.
Earlier this year, the American Medical Association published a report designed to assess the implications as artificial intelligence continues to expand rapidly across healthcare. It sought to get a handle on some central must-haves to ensure AI in practice is usable, useful and safe.
“Ensuring the appropriate implementation of AI in healthcare will require that stakeholders forthrightly address challenges in the design, evaluation, implementation and oversight of AI systems,” the report noted.
Probably the most fundamental challenge on the way to effectively deployed AI and machine learning boils down to one word: data.
“Eighty percent of machine learning is acquiring and cleansing data,” Tyler Downs, chief technology officer at MedeAnalytics, said earlier this year. “We tell our customers, that’s where it starts.”
“What’s critical to a lot of these machine learning and AI devices is the information that’s given to the algorithms to make them smarter and train them,” echoed Joseph Guagliardo, a partner at Pepper Hamilton LLP who specializes in emerging technologies. “It’s not just about the algorithms, it’s about the data that’s feeding them.”
CIOs are key, but everyone has a role to play
More and more, hospital chief information officers are realizing that job is largely their responsibility.
As one healthcare AI expert explained recently, CIOs used to enjoy a “command and control” relationship with their technology: The IT systems, “no matter how sophisticated, could only do exactly what they are told.” But as AI and machine learning have changed the way data is digested and put to work, that relationship is now “more like a two-way conversation than a set of one-way commands.”
Thankfully, most CIOs are acclimating well to this new environment, with many of them putting in place new strategies and tools, such as dashboards, to manage AI in practice, and ensure it’s working optimally wherever it’s deployed, whether clinically or operationally.
But the number-one must-have for good AI is good data. And that means an all-hands effort: not just the CIO, but whoever – in whatever department across the health – has their hands on the relevant data at different times.
“As we move to value-based care, the use of data becomes paramount,” said David Kho, chief medical information officer and chief digital officer at Miami-based Chenmed Group, told us earlier this year for our feature about the foundational importance of data governance. He emphasized five specific roles within a given organization:
- Data stewards. These are subject matter experts from different domains who are in charge of specific data systems, lead ongoing data quality and remediation work and are responsible for putting procedures in place to ensure the security and integrity of the data.
- Data owners or sponsors. These are execs or department heads, accountable for specific datasets.
- Data custodians. Like stewards, these leaders are also often focused on enforcing business rules about data access, custody or exchange.
- Data producers. This could be a wide array of staff, on both the clinical and business sides, whose activities generate data;
- Data users. These are the folks across the enterprise who are tasked with doing analytics and deriving useful knowledge from various data sets, ideally able to access and use it all, easily and effectively.
“We have to have a shared meaning when we talk about data quality,” Kho explained. “We have to have a shared meaning when we talk about a single source of truth.”
Have a strategy, and expect hard work
Much of which is to say, essentially, that if you want machine learning to work for you, you need to work for it.
“Data sources are becoming increasingly varied, and ML and AI platforms are struggling to keep pace,” writes Paddy Padmanabhan in his recent book, The Big Unlock: Harnessing Data and Growing Digital Health Businesses in a Value-based Care Era. “Aggregating and analyzing all that data is not easy.”
There’s lots of technical advice out there from companies such as Google and Microsoft with tips on how to prepare large data sets for AI algorithms.
But the specific strategies followed by a given hospital will usually depend heavily on factors unique to their own specific needs.
“Idiosyncrasies of a healthcare system can affect the performance of AI tools in unexpected ways,” said Sujay Kakarmath, post-doctoral research fellow affiliated with Partners Connected Health and Harvard Medical School at the HIMSS Precision Medicine Summit earlier this year.
What’s common across the board, however, is the need for a data management strategy – one that’s well-considered by stakeholders across the enterprise – that aligns with the goals it hopes to accomplish.
That may be supply chain or medication management or imaging analytics or precision medicine. Or all of those and more. But the data that fuels that augmented intelligence needs to be intact, complete, accurate and well-groomed if any of it is going work.
And leave aside any questions about financial improvements or operational efficiencies for now. Sound data is needed for safe and effective care in the most mundane of times – let alone with the added wild card of AI.
Richard Staynings, Clearwater Compliance chief security and trust officer and member of the HIMSS Privacy and Cybersecurity Committee explains, “the creeping influence of artificial intelligence in healthcare, where human decision-making is more and more removed from the game” means that, if data is missing or inaccurate, “the quality of patient care declines significantly” and safety concerns could arise.
At this year’s HIMSS Precision Medicine Summit, several clinical experts offered their own perspectives on the need for a smart data strategy for AI.
“At the provider level, one of the things we’re trying to figure out is how much data science do you need to teach a medical student to be able to talk to a data scientist?” said Adam Dicker, chair of Department of Radiology Oncology at Philadelphia’s Thomas Jefferson University.
“For a lot of the algorithms that underlie machine learning, there’s a fair bit of statistics,” he explained. “And for those people who want to get into the weeds a little bit, if you don’t understand the process, you really can be bamboozled.”
But getting into the weeds with data of all shapes and sizes is a necessary prerequisite for AI success, said Richard Milani, MD, chief clinical transformation officer at Ochsner Health System.
For all the hype about AI, even when there’s just as much well-warranted optimism about what it can do for healthcare, “we have to have a level of scientific rigor before we just start throwing these things out to the population,” he said. “We just have to do our due diligence in terms of making sure this has been validated in multiple systems.”