Leveraging AI for healthcare: 3 questions with digital health innovators Austin Ogilvie & Ash Damle

Jessica Kim Cohen -

Healthcare organizations are increasingly focused on integrating predictive analytics into their workflows — but as this trend continues to grow, what barriers are standing in their way?

For many healthcare innovators, the challenge is translating advanced analytics into a user-facing application. The algorithms data scientists use for their investigations are often written in coding languages like R and Python, while digital applications are built in Java — creating a rub for those trying to bring the two sectors together.

San Mateo, Calif.-based startup Lumiata faced this issue when working to deploy its health risk assessment tool, called the Risk Matrix, which leverages AI for personalized medicine. To bridge the gap between the AI algorithms created by its analysts and the digital applications created by its developers, the company partnered with Yhat, a Brooklyn, N.Y.-based software company that uses machine learning to solve coding incompatibilities.

Ash Damle, CEO of Lumiata, and Austin Ogilvie, CEO of Yhat, spoke with Becker's Hospital Review about the challenges and opportunities they faced when translating advanced analytics for the consumer population.

Question: How do these coding incompatibilities between AI algorithms and digital applications arise?

Austin Ogilvie: You can think about the data science lifecycle as starting on the far left, with raw data and business problems, and ending on the far right, with polished user-facing products. A very basic example would be the Netflix recommendation system; there's raw data that goes into predicting a user's preferences, but that's not visible.

There's a lot of work that needs to be done to bring the raw data side together with artificial intelligence and machine learning techniques, in an effort to produce a product that can be used by a normal person — one who doesn't sit around thinking about artificial intelligence and machine learning all day.

On the operations side, you have statistical programming tools like R or Python, which allow for data scientists to build based on business logic, with very advanced machine learning and AI techniques; at the same time, application developers tend to build with the mobile and web application tools that they're more familiar with. Languages like R and Python are very complicated to reengineer into these user-facing application languages, not to mention, rewriting the codes is time consuming, error prone and expensive. By bridging the gap, analysts and developers can both work with the tools that work best in their own environments.

Q: What additional challenges do companies developing predictive healthcare tools face?

Ash Damle: There are a core set of data challenges in healthcare, the first of which is AI-enablement of data. Many people and organizations have data, but it might not be the most clean or structured — the question becomes figuring out how to bring that data together in the right way, where you can structure it, normalize it and enrich it, so that you can truly take advantage of the AI that you're building.

That's a core problem, but even on top of the initial data problem, there are the challenges of actually understanding the AI techniques, understanding how to validate and understanding how to implement. A challenge is figuring out how to get real-world results: How do you validate your models in a way that actually mimics the real world? How do you ensure you're getting the performance you're looking for?

Another challenge is, when developing and deploying the predictive tool, how do you package the results into someone's workflow — simplifying the important parts, to facilitate its entrance into the real world? You need to format the output in a way that engages people, all the way from the analysts to the physicians.

AO: When you look at something like, again, the Netflix recommender — or any sort of recommendation system — the stakes are relatively low if you get it wrong. On the other hand, if you get it wrong in healthcare, the stakes are extraordinarily high.

While there's a lot of money on the table in terms of how hospitals and physicians and research labs in healthcare can benefit from these technologies, it's not nearly as easy to experiment, and it's not nearly as easy to pursue from a compliance and regulatory perspective.

Thematically speaking, healthcare applications with predictive analytics or artificial intelligence are not going to be completely automated; they're going to be bound to decision support systems. We need to realize the strongest areas to apply these tools or techniques is not in pure 100 percent automation, but more in terms of building systems that enable human beings on the front lines to do more than they otherwise could.

Q: AI and predictive analytics are increasingly being applied to the healthcare sector. Looking to 2017, what are some key areas for growth?

AD: When we think about all of healthcare, we're always trying to figure out where someone's health is today, how their health is evolving and what we can do to maximize their best possible outcome. AI has an extraordinary opportunity in healthcare, because it can help us learn, so that each new person is not an experiment — we can learn from the events and actions of all people across the board, and use those insights to support each individual.

I think we're just at the start of what AI can do in terms of healthcare, all the way from insurance to how we understand and get ahead in disease prevention. But when we think about risk stratification, risk adjustment, quality maximization and other processes that already exist today, these can be scaled much more effectively by combining man and machine.

What's fascinating is that, when you pair a human being with the AI, you can almost always beat each one alone. The outcome is just so much higher.

Copyright © 2024 Becker's Healthcare. All Rights Reserved. Privacy Policy. Cookie Policy. Linking and Reprinting Policy.