Luis Louro - Fotolia
Heterogeneous data sharing in healthcare vital to personalized care
In this Q&A, Life Image CTO Janak Joshi talks about heterogeneous data sets and their importance to healthcare AI companies, and makes a pitch for imaging data to ONC.
A seamless exchange of health data is a win for patients, but it would also be a win for researchers and clinicians who'd have better access to heterogeneous data sets.
Unlike homogeneous data sets, heterogeneous data sets are highly variable. That kind of variability is necessary to build robust data models that can provide more precise and more personalized care, according to Janak Joshi, CTO at Life Image Inc.
In this interview, Joshi discusses why heterogeneous data sets are so important to healthcare companies -- especially those building AI products, and why Life Image, one of the largest networks for exchanging image and clinical data, hopes CMS and ONC will look to the company as a blueprint toward interoperability.
Why is interoperability in healthcare such a big topic right now?
Janak Joshi: Interoperability is not a new topic. However, more recently, some of the regulations from ONC and CMS have been cross industry as far as being able to share the data, being able to manage the patient consent if you're recruiting patients for a clinical trial. And, oh by the way, you have to include a lot more than just claims data for running trials if a pharma company is partnering with a network of hospitals or if a pharma company is either considering an acquisition of a healthcare IT company or partnering with a healthcare IT company.
Essentially, it's a continuation of the maturity of the industry in the interop space in lieu of Roche Pharmaceuticals buying Flatiron Health and Foundation Medicine [in 2018] and the growing interest from pharma companies to either partner with or absolutely acquire a company like AthenaHealth and Allscripts.
The question remains: How does the convergence of broader industry constituents work together in a way that addresses the governance, the management, and the overall access component of the information flow coming from various instrumentation data sets and various healthcare IT systems?
At the end of the day, you don't really care if information is coming from Epic or Cerner or Allscripts or GE or Siemens. You care about, for example: What are the different clinical pathways? What are the most effective ways to address an intervention for the best outcomes? What is the variability in treatment patterns across different institutions for patients with similar conditions?
You can't really answer any of those questions unless you have a heterogeneous data set.
What is a heterogeneous data set?
Joshi: In order to really train your model to be vendor agnostic, you need three things: You need multivendor data, you need multi-institution data, and you need a multidemographic data.
Joshi: If you're an AI company that has FDA approval, that has really sophisticated algorithms defined by a high level of accuracy, specificity and sensitivity of the statistical algorithm yield, that's all fine and dandy. But a major challenge for an AI company and a machine learning company is the training and validation on heterogeneous data sets. Because today, in the imaging space, there isn't a large, broad data set representative of, A., multimodal, and B., a heterogeneous data set big enough that any vendor can claim success behind the model.
Do AI companies face other challenges in the healthcare space?
Joshi: A big part of machine learning is around continuous validation or self-learning. How you can deploy and scale your business model as an AI company beyond a few pilot sites? Are you really going to negotiate contracts with 100, 200, 500 hospitals? Probably not, because you can't scale a business that way.
I would say almost 75% of all of our existing contracts and proposals in the AI space are associated with distribution because these companies have spent 18 to 24 months of cycle time trying to negotiate contracts with hospitals unsuccessfully.
It's not that there's a lack of willingness, a lack of incentive or lack of clarity or ambition on the part of the hospital administration to adopt these technologies. The challenge is more logistical: For a hospital CIO, who is the buyer today for AI and machine learning technologies, how can he or she find 50 different contracts, get into absolute integration hell with Epic and Cerner across 50 different companies, and get into an upgrade nightmare where each one of these AI companies is on a different stack, a different workflow, and consumes information in a different way.
It's a security nightmare. No hospital CIO on the planet will tell you with a straight face that this is great and I'm going to do this.
Does Life Image want ONC to look at how the company operates and use this as a model to get to interoperability?
Joshi: In a best practice scenario, yes. I think more than a model -- at least as a blueprint/framework to allow ONC to recognize that imaging and genetic data is an important constituent in the broader care delivery framework. And we want ONC to also recognize that you cannot ignore imaging, which represents more than one-third of all Medicare and Medicaid payments associated with the clinical care. Imaging is the primary clinical endpoint when you talk about interop and specifically in context with the development, variability, cost containment and benefit management.
Editor's note: This Q&A has been edited for clarity and brevity.