December 11, 2019
Top Five Podcast: Janak Joshi, SVP & CTO of Life Image
In this episode of “Top Five,” we catch up with Janak Joshi, SVP and CTO of Life Image. Top Five is a podcast series recorded at RSNA.
Q: Janak, AI has been the buzz at RSNA for several years now. Taking a step back, where are we? What is the state of affairs in terms of AI in the medical field?
A: It’s a good question and the right question because today in the market everybody feels that AI is going to solve a lot of issues with regards to accessibility, variability, and even the radiologists’ ability to augment and make better decisions. I’m taking a skeptical route to that approach in the market because, for me, AI is where the Internet industry used to be back in 1997. So I think we’re in the 1997 of AI for a variety of different reasons; the biggest reason being the validation behind these AI models. If you talk to the hundred plus AI companies and ask them “what data set did you train your model on,” “did somebody validate that model,” and most importantly, “was that training data heterogenous,” which means representative of multiple hospital systems in multiple vendor systems. The short answer is no, it was not. Most of these AI vendors have trained on publicly available data sets, which are 10 plus years old and are probably just coming from a single vendor, like GE, Philips, Siemens, etc. When you make a claim that “I trained an FDA approved model on a hundred patients, using ten year old data, coming from a single vendor system, in a single institution,” I got a problem with that because I don’t want to be diagnosed or, at least my diagnosis be relied on instruments, in this case the AI being the instrument, that have limited accuracy and specificity and sensitivity based on the training data out there. The other big issue with AI models is AI is still a tool. As a tool, you need shelf space to distribute it. Distributing it causes awareness, awareness causes adoption and that is a fundamental operating piece that’s missing from most of these AI companies. Most of these AI companies are not plugged into networks like Life Image, Quest or Illumina, which means that if you have an AI to detect breast cancer they should technically be talking to Life Image because we own the biggest breast cancer imaging traffic in the United States. The fact that they are a) not aware that they need to distribute the AI and b) every AI company feels that they’re going to build their own network, going to make their own individual sale to a hospital system is very unrealistic. That’s exactly what happened between 97 and 2000 before the Internet crash.
Q: How do we break through that hubris?
A: I think the perception and the notion from the tech community, both in the Bay Area, as well as in Boston, is that you can throw a hundred million dollars at building an AI company, get FDA approval, get the European Union approval and things are going to be great is unrealistic. There needs to be a tremendous amount of education from RSNA, from HIMSS and from other peer-reviewed organizations, such as ACR. Specifically, educating that adoption is not based just on accuracy. Adoption is based on a pattern. Adoption is based on feedback coming back from radiologists and physicians and patients. That piece is missing today, number one. Number two, I think we as network vendors and network operators need to increase the velocity, frequency, and veracity of educating the broader industry, both in the provider, as well at the pharma space, and indicate this is how AI distribution is going to work. Most importantly, I would say we need a unified approach on the validation exercise. There is no central body which is peer reviewed that indicates that if you’re detecting cancer at rest that this is the data set you need, this is the phenotypic data that you need, and these are genetic biomarkers you need to consider before you make a diagnostic claim. That body does not exist today. Now, whether it’s a non-profit driven by RSNA or whether it’s in collaboration with the industry we will have to figure it out. The lack there of is causing further fragmentation, confusion, and it’s potentially contributing to patient safety events, which people should be aware of.
Q: Discuss Life Image’s solution set to help validate and provide more robust, heterogeneous data to AI companies, in coupled with the network model.
A: If you break down the core components of what makes an AI company successful, the notion that accuracy is what AI makes a successful business model, according to me, is not right. I think what makes any company successful is adoption and continuous validation based on real-world evidence datasets. Meaning, that when you think about feeding an AI with data to train the model it’s not a point-in-time solution, rather, it needs to be a living data set that continuously feeds the model from real-world evidence across heterogeneous partners, vendors, as well as hospital systems, to kind of train that model to a point where it’s industrialized. Currently, it’s not industrialized. That translates into three very distinct capabilities that Life Image is bringing to the table. Number one is the data asset. Over the last several years Life Image has been working hard at bringing novel data assets to the market, both for pharma companies, as well as for research informatics, as well as for AI companies to mature the evidence lifecycle. Number two, Life Image is a network, which means that we have the plumbing and the pipes across all the clinical endpoints, across the workflows, whether its primary care, whether its specialty, or whether it’s inpatient, whether it’s ED, whether it’s independent imaging centers, which means that if you’re an AI company and if you want to scale to a hundred hospitals, a thousand hospitals, two thousand hospitals, Life Image is able to facilitate that today, based on our distribution process and distribution framework for AI companies. We are already doing that across almost half a dozen AI companies, stroke being one of them, mammograms breast cancer detection being the second one and so on. The third and the most important piece where we are assisting AI companies is the validation piece. Most of the questions that we are getting from the industry today is around, “do you have heterogeneous data for us to make, improve, or mature our arguments with the FDA” or “I’m already an FDA approved organization as an AI company and I’m trying to essentially extend my adoption and provide better arguments for the medical community by virtue of which I need more representative data to train the model.” Those are three core areas where Life Image is already engaged, both in the U.S., as well as in Europe.