AI and Healthcare: Diagnostics

Interview Transcript

Article | AI and Healthcare: Diagnostics
5th June 2020 Atheneum Team

Expert Profile

Role:

Vice President of Sales

Organization:

Oure.ai

Bio:

John Hawkins has over 30 years’ experience in healthcare technology and strategy, including 12 years’ clinical experience. He has worked in consulting, business development, sales, and IT across a wide range of healthcare networks and technology companies. He is currently employed as the Vice President of Sales at Qure.ai, a US-based AI startup that improves patient care by leveraging AI and machine learning.

Section 1: The Basis

1.1. How has the uptake of AI for diagnostics changed over the past 2-3 years?

It has certainly increased.

Now, before where we might have focused on understanding the “black box” of information – i.e. how X piece of information got to the output – what is of greater importance now for clinicians is trust. Trust comes through variation, validation, and verification. We call them the three V’s of artificial intelligence.

Firstly, you train your algorithm, for which you need a lot of variation in your dataset. Variation helps to minimize bias. As in the words of Mark Twain’s, “If all you have is a hammer then everything in the world is going to look like a nail to you.” Variation, therefore, remains very important. You need to show variation in terms of patient demographics – old, young, male, female, different geographies, different ethnicities. This also applies to the types of machines that produce images.

Secondly, you need validation. You have to ask yourself, “how are you validating your algorithms?” It is one thing to say, “we are at 99% accuracy in detecting a cardiomegaly or 98% accuracy in detecting atelectasis”. However, until you have published studies and pushed those out through peer review, and a regulatory body like the CE or the FDA has looked at and validated that your solution does what it says it’s intended to do, you’re not going to get very far.

Thirdly, you have to verify the results, which is where the radiologist comes in.

Ultimately, people are realising that it is less important to understand how the algorithms work. There are lots of diagrams out there now about how this all works, and it is pretty well-documented – I call it “algebra on steroids”. All the calculations focus on the probability for the area under the curve, i.e. the accuracy. It is less important to do that and more important to have trust in the validation of the solution and as this trust in validation has increased over the past few years, so too has usage increased.

1.2. Why is diagnostics such a target area for AI-applications?

In terms of why AI is becoming the nouveau thing for diagnostics, the first thing you have to remember is that medicine is both an art and a science. At the end of the day, the doctor is trying to diagnose, prescribe, and treat. Oftentimes, they get it right but, more often than not, they get it wrong. This is because doctors are humans and humans are not infallible – they are prone to mistakes just the rest of us. They might be tired, overworked, or just disengaged. Any or all of these can factor into a mistake. If AI usage can reduce potential mistakes, then that is desirable.

Indeed, the precedent for computerization was set within the laboratory. 25 years ago, in an environment where there was a history of pathology, you would have a group of trained lab techs prepare a slide. They would take a blood slide, smear it, put it on the slide, but that under a microscope, and then they had to count white blood cells, red blood cells and other things they saw with a clicker: click, click, click, click, click. The lab companies pioneered computer diagnosis through needing to do the counting and they were able to automate the processes that the lab tech was used to doing. Initially, you must understand, the pathologists and even the clinicians rejected the developments as they didn’t trust them. It certainly took a little while for the validation studies to be published and for trust to be built between the machine and the clinician. Now, however, it’s a lab standard of care.

AI is a particularly compelling solution specifically for radiology diagnostics because almost all of radiology today is digital. Granted there are differences as it might be for a CAT scan, or an MRI or a PET scan, or even a plain film radiograph. The common thread, though, is that it is all digitized and because it’s digitized, we are talking just about the analysis of pixels. The trained machine is much better at diagnosing the pixel level than the human eye. The machine can read 1,000 x-rays a day, where the human eye might only be able to read accurately, perhaps, 300 x-rays in a day. The benefit is found in getting the machine to do the dull, duplicative work that the human would typically be charged with. Now, the paradox here is that the machine can’t make the final diagnosis, which is why the radiologist has to be there. I recognized this in 1991 that the computer should be augmenting the role of the radiologist and not replacing them.

To answer your question more explicitly, radiology is a digital practice built on image recognition and classification in terms of algorithms. Why not let the machine either recognize all of the studies before triaging the normal from abnormal studies? With this in place, the radiologist comes to work every morning just like anyone else does when starting their day: they come into work, turn on their computer, and work through their inbox. Just like anyone else, they look at their mail and try to filter out which mail is most important, what is junk or social media mail, what is high priority mail versus ads, etc.

1.3. How has the trust of AI changed/continue to change?

I think you could almost do an age-based stratification of clinicians in trying to answer this question. I have had conversations with older radiologists and the first question they have asked me is “How accurate is your solution?”

We respond by telling them that we are at about 97-98% accuracy. The older radiologists will get all defensive and say, “How can I trust your solution if you’re not 100% accurate?” I’ve often scratched my head at this response and have asked a few doctors about this to challenge them. I asked them, “Well, doctor, in the best minute of the best hour of your best day, are you ever 100% accurate?” The answer is, of course, no, they’re not.

Part of the validation studies that are performed will compare an algorithm to a panel of radiologists and the goal is to demonstrate that the solution is as good as – if not better – than that specific type of radiologist in identifying certain things. This is possible because radiology is self-assessing. There are often A-Readers (the primary reader) and B-Readers, who might come in and conduct an overread for that A-Reader. Even then, the error rate remains reasonably high and accuracy pretty low in most radiology departments. To add a bit more specificity to the question of why AI is good for diagnostics, you see that it in radiology functions as a great B-Reader. Consider an example, in which the radiologist is in a hurry, is tired, or maybe his eyes are sore because he was yachting on Sunday, and what AI offers is a “security blanket” for them. I have heard radiologists describe it as such.

If you think about it for a moment, you can see the value. For example, as a radiologist, I might read a CAT scan or an MRI, both of which might have 140-180 slices – that is 140-180 slices per study for just one patient. Then, they might need to compare that to that same patient’s previous study and they’re trying to annotate the delta or the change. Consequently, having the computer function as the B-Reader helps the radiologist sleep easier at night. Once the radiologist realizes the potential of using AI as a B-Reader and has a bit of experience with it, the level of trust increases. This has taken a bit longer with some older radiologists, but we are starting to change occurring.

The other important aspect to consider is, “how do radiologists learn to trust the AI?” As with robotics, if the work is dull or duplicative, AI can do a great job in measuring and assessing the change of a condition throughout a given period of studies. Take the brain, for example. As we know, the brain has two hemispheres, the left, and the right hemisphere. However, there is also something called the midline, which sits between the two hemispheres. I have sat in reading rooms and watched a radiologist reading a neurology case where there has been a midline shift (a mass or a bleed that has forced the brain shift to one side) and pull out a pair of little calipers and a measuring tape in an attempt to carry out a manual measurement of a midline shift. They might also do this when trying to measure the volume of a mass because they want to compare that to the previous study – you often see them scribbling on paper what the changes in volume are. This brings us to the other great use case for AI in quantifying and measuring the progression of a disease. The AI is brilliant at this and can do it much faster, much more accurately, and much more consistently across the board than the human can. As you can imagine, this begins to engender trust with the radiologist once they realize this is the case.

Now, depending on the reporting format, the increased accuracy can also improve radiologists’ ability to satisfy their stakeholders better. This allows the radiologist to more accurately communicate to the neurosurgeon what the midline shift was at the start of this disease. As a neurologist, you can subsequently tap the brain to relieve the pressure in a specific area with greater confidence, causing the midline to shift back to its normal area. To show that graphically is huge. The same thing applies when there is a mass in the brain. Before treatment, the mass was 4cm by 3cm. After treatment, the mass now is 0.5cm by 1cm. To be able to report that graphically on the radiology report directly improves stakeholder management with that internal customer – the neurologist or the neurosurgeon. If a solution can improve internal stakeholder relationships, that radiologists will positively be disposed to use and trust the solution.

1.4. Which medical disciplines are leading the way? Why?

I do think radiology is leading the way, although more companies are now creating solutions for other disciplines.

IVF is a good example. In this case, eggs are harvested, sperm and eggs meet, and now you have an embryo. What typically happens is that you have a lab tech who’s looking at a platelet or a petri dish of growing embryos and they are trying to select, based on certain parameters, which embryo has the best chance of survival. The whole process is very manual and labour-intensive. There are a couple of companies now that are looking at tackling that using image recognition. They have created algorithms to define what the healthiest embryo looks like, and then they are using AI to select those embryos.

The other interesting thing from a non-clinical, back-office perspective in healthcare, specifically in America where the insurance rules and regulations are so convoluted, is that we are starting to see companies create algorithms to optimize coding and billing for receivables. We are starting to see AI being applied to the workflows in the back-office billing side. That’s going to be an important leap forward once it becomes more widely utilized.

I know there are some other areas in healthcare that are adopting AI as well. Obviously, in biotechnology and pharma research. Rather than running what we call a Monte Carlo simulation, the computer can just generate tons and tons of data based on simulation alone. So, you’re seeing AI penetrate the biotech space too, especially right now with COVID-19.

Section 2: Market Dynamics

2.1. How are companies like Qure.ai differentiating themselves?

I think you must understand that market direction will always be driven by value. 90% of AI’s dirty little secrets rest in these two areas. The first is, “where is the value?” In other words, how does applying an algorithm or AI solution add value to the diagnostic and treatment workflow?

Qure is starting to get clarity on that within radiology. It could come in the form of pre-reading all the studies and triaging the normals from the abnormals. Workflow enhancements could result from either or both of the pre-read/post-read by making sure the radiologist did not miss anything. Adding value to the workflow is dirty little secret number one that you need to understand to differentiate yourself. Ask yourself where, why, and how your product is adding value to the workflow.

Dirty little secret number two is you have to have someone to pay for it. Business models in this space are constantly evolving and this means that there is a wide range of price points on the market, which can be challenging to get payers on-board. Moreover, a lot of AI companies are saying, “Do a pilot study!” or “Try our product for a while, you’ll love it!” The radiologists will, of course, try it out but then they will say, “So what? I still can’t afford to pay for it because my margins are too small”. The issue with margins in the USA is true irrespective of whether you are talking about a for-profit, not-for-profit, or government-funded health care facility. Ultimately, margins are so slim that nobody wants to pay for it. If you can get someone to pay for it, then you will have the opportunity to differentiate yourself, which is something that not everyone can get. That is dirty little secret number two.

Given that we are still in our infancy in applying AI from a clinical workflow perspective, how do you go about getting traction? Firstly, you have to solve the two dirty little secrets: you have to solve the value issue and understand how you are adding value and then find someone to pay for it. How do you get there? You get there by demonstrating your solution adds so much value so that becomes a clinical standard of care. If you can get there, then you have a better chance of finding someone to pay for it. I think companies like Qure understand this…not that it makes the process any easier.

2.2. What opportunities for collaboration/M&A are likely to emerge?

Within the US – and likely globally, as well – once a solution becomes a clinical standard of care, then you have to win over the lobbying groups. You need to get the American College of Radiology or the American College of Pathology to support and write position papers and petition the government Centres for Medicaid Medicare Services to create a reimbursement code. Once there’s a reimbursement code, then the insurance companies or the government is obliged to reimburse those organizations that have adopted and implemented that solution, baking it into the clinical standard of care pathways. At this stage, there are far more opportunities for collaboration because you are established within the market.

There are a couple of precedents that have been set for that one in the mammography space, in which computer-aided diagnosis emerged around 12 years ago. Early market leaders relied on using CAD although, initially, radiologists rejected these solutions. Over time, however, through validation and verification, it has become a standard of care. This solution now has a reimbursement code, which opens the door for lots of collaboration opportunities.

There are some similar precedents in the MRI of the brain for central nervous system diseases. Although 3D rendering of the brain was initially rejected by radiologists, enough companies got traction and deployed a solution that got sufficient attention in the radiology world to receive a reimbursement code for the 3D rendering of the brain. Now you have that conversation going with the radiology department who asks to use a specific solution to make them more efficient, the result is that they are working faster and spending less time with their calipers measuring. As a result, they are more effective, which means that the reports are sexier for the neurologist and the neurosurgeon.

Solving a specific problem is going to be one of the biggest momentum gainers that will prove potential and lead to collaboration and perhaps M&A opportunities. You also have to have a vested stakeholder community to continue moving the ball forward, of course. In summary, opportunities for collaboration are discipline-specific and are closely tied to exposure.

2.3. How might some of the more established tech giants look to get involved in the space?

Licenses are kind of, interesting, in this context. There are specific business models tied to end-user licenses (EULA), while other AI companies are sub-licensing themselves to OEMs. Take GE, for example. GE manufacturers an enormous amount and has a huge portfolio of diagnostic capabilities, whether in radiology or cardiology. The real question that OEMs like GE and others have to come to grips with is whether buy, build, or lease the solution?

I think OEMs are going to start looking at a lot of lease models for the next one or two years to build those software capabilities into their product. They’ve been doing that successfully for years, whether it’s work scheduling solution or licensing a view box or an imaging view box. So, large OEMs sub-license different components already, and they’ll be sub-licensing AI as well.

Another point to make is that there is a tight relationship between the biotech and the pharma world, clinical research, clinical trials, and leveraging AI for data management within the provider world. At the end of the day, biotech and pharma work through clinical trials, and they need diagnostic centres and they need patients. All of this generates an enormous volume of data. AI is a great solution for navigating that data. So, biotech and pharma companies can embrace that by aligning themselves to these solutions, identifying knowledge leaders, product leaders, software, and service companies that do something well, then baking that software into their clinical trial formula.

Right now, a crucial thing to note in terms of dynamic market changes is consolidation. There is no question in my mind that the market will consolidate. Until about February 15 this year, I would have said that the bubble was growing. This is because there were lots of new startup companies coming onto the scene. There was a lot of hype within the industry, to be honest. Some notable companies were cloud-based or SaaS-based and had everything laid out. However, with COVID-19, I think what we will find over the next 12 months is that the bubble will burst a little bit and some leading companies will emerge. I think perhaps 12, 18, or even 24 months from now, we will start to see real consolidation occurring.

I think this is already starting to happen through AI marketplaces. There’s a great company called Blackford, which is an AI marketplace. They are just tagging all kinds of AI companies that offer different modalities and different solutions. Entry criteria onto the marketplace platform are that the solutions need to be FDA and CE certified. Blackford then turns around and offers those solutions to its customers. You’re starting to see that type of consolidation as well in, not only OEMs but more broadly in software.

I think we can all see some potential here for large tech companies to get involved. Whether they will or not remains to be seen.

2.4. What are your thoughts on the dos and don’ts of bringing an AI-driven diagnostics solution to the market?

Your go-to-market strategy must include the following:

Firstly, having a product that’s been validated and stamped by either CE or FDA. If you’re not validated, then you’re going to spend a lot of time proving yourself. So, validation is the critical number one.

Secondly, I’m going to highlight what we are doing right now in this discussion, which is building content. You must build and push content – you must generate buzz. You generate buzz by having notable experts and key opinion leaders championing you, using your solution in their workflow, and baking your solution into that workflow as a standard of care so that others can say, “Well, Dr. Jones in Philadelphia is using it and so I should try it too.”

The third “do” is what I shall call “flawless execution across the board”. Flawless execution is used here in terms of integrating, not only in the technical ecosystem but in all the systems you need to share data with. This includes:

• Understanding the sales cycle;
• Understanding who the economic buyers are;
• Understanding who the clinical decision-makers and technical influences are and getting them on-board;
• Understanding the procurement cycle at the organization looks like and whether the target organization is government-funded or private;
• Understanding whether there is an RFI or an RFQ in place or if the transaction is merely a sole sale;
• Understanding the specific workflow of the target discipline;
• Having the existing clinical credibility to suggest where the adoption and touchpoints are at the established workflow, and then having the training and the capability to hold the hand during implementation and post-implementation and training; and
• Flawless execution in terms of delighting the customer. Client satisfaction, so that that customer turns into a large champion for you.

You must remember that a lot of these AI companies have brilliant data scientists and not clinicians on their teams. Fundamentally, they are brilliant engineers who love to create solutions. Smart engineers will identify a problem to create the solution for, but a lot of AI companies right now are just brilliant engineers that have created great solutions and now they’re running around looking for the right problem to solve.

I think, in summary, implementation starts from having a deep understanding, winning over key opinion leaders, achieving technical integration, holding the hand of the client, and then ensuring client satisfaction.

Section 3: Implementation and Changes

3.1. What barriers to implementation continue to exist?

Another barrier people often mistakenly this exists here is licensing. However, licensing is easy. Essentially it is just, “Hey, we’re going to license your software!”

The business model itself, though, is where you must focus as it can be closely tied to the issue of implementation. This is because the answer to “How do you pay for that software” might still be a work in progress. Some AI companies are offering their product as a pay per click to offset this. A client agrees to use their software for one study and then pays for one study.

The other use cases rely on a software license fee structure based on either the number of users or on volume. For the former, you might have 20 people using the software for the next year, and so you can bill accordingly. For the latter, you are going to use this solution for 500,000 studies, which has X price associated with it. If your volumes are higher, then we’ll lower the price – the typical volume-based discount.

Having said all of that, if we think about the industry now, the three key barriers are workflow, workflow, and workflow. The key gets back to the workflow. How do the clinicians want to see this solution adding value to their workflow, right?

Then, that last little thing about implementation comes down to cost. How much is it and who’s paying for it? But, getting the clinical champion that says, hey, I need to use this in my diagnosis or treatment, the standard of care is going to be critical.

3.2. How does implementation differ geographically? I.e. US vs ROW?

We are seeing an increase in AI use in third world countries, driven by a shortage of competent clinicians. Those markets are ripe for telemedicine, remote monitoring, and artificial intelligence. One of the use cases that CAD4TB or Qure is that they have a great chest product that’s important for screening for tuberculosis. TB is a huge public health concern and is on the rise. It’s a disease we should have resolved years ago, but it’s still on the rise. The lack of a radiologist could delay the diagnosis of TB by three to four weeks. Applying radiology at the screening clinic can identify TB cases within minutes, not weeks. So, once you’ve identified a person is susceptible and demonstrates a disease in the study, you can then isolate the person and treat them using the correct protocol.

A lot of companies are now responding to third world countries like Pakistan and India, and South America, or South Africa, through chest x-rays just to look for ground-glass opacities. I think, in summary, the third world countries are just a petri dish for AI solutions.

But, guess what? Where is the funding going to come from? Dirty little secret two rears its head again. Some examples of progressive funding bodies include Stop TB and the Gates Foundation. We are seeing a lot of large NGOs funding some of the use cases and within the space, some companies are carving out some commercially viable products…but none of them are a Tesla, that’s for sure.

3.3. Data privacy is a key concern for many patients. What challenges exist in ensuring compliance for patient data for AI and ML applications?

Being able to comply with GDPR and PHI and HIPAA in America is critical. Having really smart lawyers to craft your license agreements is essential – these are number one and number two. What I think we are seeing is companies are just shifting the blame upstream. Often, it is kind of a cop-out. However, you can only really push privacy laws upstream through having great software license agreements and ensuring that the ecosystem that’s adopting the AI is, itself, applying good GDPR. Part of the sales cycle is understanding the IT requirements of the client and demonstrably ensuring compliance with IT requirements and security requirements within that cycle.

I think the bigger issue here to discuss, though, is liability. When thinking about AI use in a clinical field, especially for diagnostics, who is liable if the AI either misdiagnosed through either overreading or under reading something? These are the bigger questions that companies have to understand. I think, right now, most organizations carry liability insurance, physician malpractice insurance, etc. In reality, doctors don’t get it right 100% of the time. They probably get it wrong more than they get it right, surprisingly. As a result, doctors and institutions are liable and for good reason. Having smart end-user license agreements in place that defines who’s liable if something has been misdiagnosed, is critical.

Section 4: Outlook

4.1. What impact has/will COVID-19 have on the use of AI for diagnostics?

As I mentioned earlier, we are seeing AI penetrating the biotech space through simulation. It’s kind of, off-topic, but it makes me frightened to think about how potential vaccinations might be fast-tracked and then deployed due to COVID-19. What’s the real safety and efficacy of that? I think that is something to watch.

COVID-19 is also interesting because I know several companies are now providing a solution that looks for COVID-19 on a plain film x-ray or a CT. We’ve talked about the global shortage of COVID-19 testing kits, whether it’s the lack of the swab, or the medium to transport or the actual test itself.

4.2. What other areas adjacent to AI are being explored and may have potential cross-over benefits in the future for diagnostics? I.e. augmented reality / real-time integration.

Visage is a very large PACS (picture archiving and communication system) company with an install base in hundreds of hospitals from a PACS perspective and Visage is starting to package AI solutions as a product offering to their customers. Why? Because their customers are asking for it. What’s the benefit? The benefits to the AI company are that Visage has already gone through the contracting and all the IT and the procurement hurdles out of the install base and it, very quickly, becomes just an additional add on to that contract.

Then, I think, the last thing to note there is Nuance. Look at Nuance software. I think 80% of all radiology dictation is through Nuance Dragon Speak software. So, Nuance just created an AI portfolio as well of great companies that they’re offering to their customers. They’re still trying to answer that question of where does it add value in the workflow? Is it going to be a pre-read, a post-read? Wherever the solution sits within the workflow, the question of “Where is it adding value?” remains.