HIStalk Interviews Kyle Silvestro, CEO, SyTrue

Kyle Silvestro is founder, president, and CEO of SyTrue of Stateline, NV.

Tell me about yourself and the company.

I’ve been in the world of clinical natural language processing for almost the last 17 years. I started SyTrue a little over nine years ago, thinking about how to solve challenges around archaic workflows where we still require humans to read medical documents, especially from the perspective of CMS. And, how we automate a number of processes by eliminating inefficiencies within the system.

How has the need and the ability to automatically extract information from medical records changed over the years?

The need has been there for more than a decade. The awareness is coming to the forefront. We are truly understanding the value in information. The advances in AI and ML have highlighted that. Most of those advancements have been more around structured data and what is possible. Looking forward, organizations are understanding the value of the unstructured clinical note that still comprises the majority of information created in healthcare today. We process more than a billion pages of these notes annually, and that’s just scratching the surface. That would be on data created within the last 12 months. The majority of information is still in this format.

It depends on where you are in the process from the point of care to the point of need. At the point of care, maybe you can get high quality data quickly, but most organizations are not. They are downstream of that information, and it’s packaged up more often than not in a PDF. It’s not even unstructured data — it’s an image. That image is shared with organizations and data is often needed 20 ways downstream. If you don’t have a way to create this exponential uplift, then you can’t start addressing the challenges we see in the system. This problem has been here for a while and there are truly no good solutions addressing it that have a critical level of adoption.

Do PDFs usually come from outside facilities, meaning that it’s an interoperability problem, or are they self-generated because the source system doesn’t capture the data discretely?

It’s a combination. More often not, this is a byproduct of a record release process. Thousands of people go on site to facilities every day to get data from hospitals or provider offices. There are some electronic exchanges now, with CCDAs being sent across the wire, but that’s really the two ways that they are getting this data. It’s definitely an interoperability issue, but it’s more of a misalignment of incentives that is potentially preventing wider adoption.

What are payers and CMS doing with the data?

We have a unique challenge within the payer market. So much of what they get is an image, a PDF that can be thousands or tens of thousands of pages long. The only answer before SyTrue was to assign a nurse to read the document, go through the 4,000 pages, and find the eight or 10 pieces of information that answer the question. But more often than not, the 5,000 other data points that are in that PDF document that could be driving exponential uplift within an enterprise are left behind. They’re saved as an image, so they are being lost. The knowledge that is in front of them is gone. Our solution addresses the efficiency challenge, but we can also liberate all of that information to drive exponential downstream value on an enterprise level, to be able to create standardization and interoperability that can drive change.

What is involved with taking a PDF document and turning it into useful information?

This is a differentiator between SyTrue and everybody else. I had the privilege, or the advantage, of failing more than most people in pushing an early technology into the marketplace. Before I started SyTrue, I implemented NLP across life sciences, payers, and providers across a number of use cases, but had also seen challenges at failure points. As somebody who doesn’t like to lose, I remember those failures. 

When we architected SyTrue, we knew that it’s not just about NLP. If healthcare data is clean, NLP is easy. It can read the document, parse it, and extract it. But the problem is that we are dealing with inconsistency from organization to organization, physician to physician, EMR to EMR. How do you account for all this dirty data that was created by a million physicians that generate billions and billions of notes annually? And if those notes are needed 20 or 30 ways downstream, you’re creating a exponential data problem that you can never throw enough humans at to solve.

That’s what we thought about. We thought about that document life cycle. We thought about the creation sources. We thought about who needs it along the way. The question that we asked ourselves is, how can we make people money along the way? How can we add value? That approach allowed us to look at it from a longitudinal perspective, because we thought that if you can get to a longitudinal data and you can do it accurately, everything else downstream becomes easy. You have all the Legos, you just have to actually assemble the house or build the car. The structural components of the information are in that longitudinal record. It’s a matter of how you are combining them. 

With HEDIS, you need problems, procedures, and HCPCS codes. Risk adjustment. You’re looking at problems and supporting conditions and payment integrity. You’re looking at elements that would roll up to make a determination — is this truly an acute kidney injury, or is this sepsis? If you have that baseline data, the downstream questions that you’re asking or the objectives that you’re looking to get out of that information become a lot easier. You can do it across many domains, as represented by our client base and use cases that they leverage.

How will the healthcare entry of big tech firms affect your business, such as Google’s work with EHR search?

How soon before they call it quits again? They’ve all taken bites of this apple, only to fail miserably. I honestly think that’s the trajectory they are on. They do the market a bigger disservice than they do a service. They push early-stage technology that’s not prime time into our marketplace. They suck the oxygen out of that marketplace, and organizations that are small and may not have the $100 million marketing budget get squeezed out. True innovation never gets bubbled up to the top because you have these massive enterprises send 14 sales reps into a client to push a product that’s half baked.

You see that in Amazon Comprehend. They just reduced their price by 95% and now it’s this big announcement around SNOMED. Great, right? If it wasn’t good to begin with, it’s not going to be better when it’s 95% discounted. We’ve had SNOMED for nine years. It’s not new. It’s not really an announcement. Talk about how you’re making people money, talk about how you’re changing the system, and don’t just make noise. That’s what a lot of these organizations do. They truly don’t understand the problem and they truly don’t understand the solution that they need to create to solve it.

IBM Watson Health had some pretty grand ambitions and failed miserably.

MD Anderson Cancer Center. The trail of tears goes on. The billions of dollars that were invested into a technology that played “Jeopardy!” and then thought it could solve healthcare was amazing. They had 5,000 people at one point. It had a lot of data. But they couldn’t roll out anything that was meaningful, except for marketing hype. That is true of many of these big tech players getting into healthcare. They don’t understand the problem that they are trying to solve. They see dollars, they think they can throw enough money to grab market share. Unfortunately, I think they do the overall marketplace a humongous disservice. I haven’t seen truly significant impact from companies that took something that was playing a video game and thought it could solve healthcare.

How do you see the investment buzz over AI playing out?

There’s real opportunity in the technology. But I think you apply technology where it makes sense. You just don’t try to brute force everything, and because there’s a new technology out, think you can solve all the problems. We take a pragmatic approach. Use technology where it makes sense to apply it. As we get downstream, AI is going to be really, really meaningful. It’s going to be important in healthcare. But we have a foundational problem today in healthcare that is going to prevent that from becoming a reality for a little while, unless organizations start to realize it. If you’re not creating an interoperable base of accurate information that you are basing your models on, you are building a house of cards. I wonder how many of those actually exist today versus true value.

There’s a lot of hype, but when you actually get into the information, what impact is it actually making? Marketing has latched onto it. Not a lot of people understand it. Everything is a supervised model. Unless you get to accurate datasets at high volume, you’re somewhat playing with fire. But we have clients that actually do this and they see significant improvements in accuracy, sensitivity, and the impact it has on an organizational level, because they are working from an accurate, interoperable piece of base-level data that’s a solid foundation.

Where will the company focus on the next few years?

SyTrue is positioned to be a dominant player across many different solutions — HEDIS, payment integrity, fraud risk and abuse, risk adjustment, social determinants, expansion of radiology, expansion in oncology — all with a single platform and with the focus of making organizations money quickly and being able to get them live fast to enable that ROI. I see great things for SyTrue. I see us going from just shy of 40 employees now to a significant number in that period of time.