Pia Abola:
Welcome to Countering COVID, a podcast by LGC Biosearch Technologies, where we speak with clinical lab scientists and industry experts to understand their perspectives on the current and future state of SARS-CoV-2 testing. My name is Pia Abola, and I'll be your host. On this episode, we'll take a closer look at the role of artificial intelligence or AI across the SARS-CoV-2 testing workflow. Here to discuss this topic, we have James Grayson.
James Grayson:
Hi, Pia.
Pia Abola:
Hi. James is an expert in the automation of molecular diagnostic workflows and a field application scientist with UgenTec, a company that provides software that automates and standardizes the entire qPCR assay and analysis workflow for molecular diagnostics. We also have Dries Hens.
Dries Hens:
Hi, Pia.
Pia Abola:
Hi. Dries is the chief medical officer and co-founder of LynxCare, a company that uses AI to help hospitals and life science companies generate insight from large volumes of data. And, lastly, we have Dirk Smeets.
Dirk Smeets:
Hi, Pia. Happy to be here.
Pia Abola:
Dirk is the chief technology officer of Icometrix, a company that uses AI for fast and objective quantitation of medical images. Very excited to have representatives across the whole workflow of SARS-CoV-2 testing outside the hospital, and then providing care once you're in the hospital. I'm going to ask you to provide a little bit of background on how your company is participating in the workflow. So, James, if we could just start with you, could you tell us a little bit about what your company is doing for SARS-CoV-2?
James Grayson:
Sure, Pia. So, UgenTec has been historically very involved in clinical data interpretation. So looking at molecular diagnostic assays, PCR, real-time PCR-based assays, and using artificial intelligence to specifically interpret the PCR curves themselves, to look at the morphology, look at the shape, determine if this is truly positive, truly negative, potentially a false positive or negative, and assign Ct values along the way. And how this has evolved for us as a company during the pandemic is that we're able to apply that technology at scale. So we've been involved in... As things progressed and over the last few months, we've been involved in large scale initiatives, some on a national level, in which in some extreme cases, we're supporting nations, companies, industries with up to over 120,000 samples analysis per day. And when you think about having to look at those curves manually from a human perspective, you need some tools to support analysis and data interpretation and results reporting at that level. And we've been able to employ that artificial intelligence to help streamline the whole automation process, to really provide scalability and decision support.
Pia Abola:
Thanks, James. Next I'd like to ask Dries to please tell us a little bit about LynxCare and how your company is helping address the pandemic.
Dries Hens:
We started LynxCare because less of 20% of medical data today is accessible for research. And the reason for this is that hospital data is either unstructured, for example, in the medical notes or scattered into different data silos. So in order to unlock 100% of available data, we've developed LynxCare an AI-driven clinical data platform, where we enable hospitals to mine their structured and unstructured hospital data, but also store them in an internationally complaint format. And by doing so, rich, granular and quality controlled clinical data becomes available, increasing our knowledge in different disease areas. And with respect to COVID-19, we developed together with leading pulmonologists and medical experts the data mining model, where currently we are extracting and aggregating over 250 relevant data points per patient in over 10 hospitals. So this database is currently one of the most granular databases available and based on our work, researchers and physicians can now analyze why certain patients declined faster than others, which patients are at high risk of needing for example, mechanical ventilation, but also what are the long-term clinical effects of a COVID-19 infection? And that's how we support as a company the hospitals involved.
Pia Abola:
Thank you. And then Dirk, could you tell us about Icometrix?
Dirk Smeets:
Sure. Icometrix is a company that is an expert in artificial intelligence on radiological images, and most specifically on radiological images of the brain. So we work in general for radiologists, neuroradiologist and their referring physicians, which are the neurologist. We do that, for example, for multiple sclerosis, where our artificial intelligence-based measures can help the radiologists to read more efficiently, but also help the neurologist to make better treatment decisions. For example, know whether a drug is given to the right patient or not. But actually, when the pandemic started in beginning of March in Europe, we thought, "Actually, this technology that we apply to these medical images can also be applied to lung images." We decided to put some efforts in building an artificial intelligence tool for analyzing chest CT scans. And we're not doing that on our own, but in close collaborations with the Universities of Leuven, Brussels, and now recently also Oxford, Maastricht, Heidelberg. And what we actually do is we look inside the chest CT scan for abnormalities that are related to COVID-19. Different patterns are occurring in the lungs of the patients. And this helps the radiologist to get a more objective view on what's happening and also for the treating physicians, it helps to decide what would be the best possible treatment path.
Pia Abola:
Artificial intelligence is fairly new to the medical profession. For applications in the medical field. And it's really not the kind of thing that clinicians or life science researchers are routinely trained in. Can you define artificial intelligence in a way that a general biomedical professional can understand? Who would like to answer first?
Dries Hens:
What is AI anyway? For me, it's just the typical definition of simulation of human intelligence processes. For me, it still is the most clear definition. And what I mean with that, with LynxCare, our algorithms just mimic manual data extraction process as accurate as possible. And by doing so, we can harness the processing power in a hospital. And for me, AI is still just a simulation of a human intelligent process.
Dirk Smeets:
I do agree with that, Dries, and what we also see when we speak with radiologists, which are our customers, is that AI is something that is not really understood. And actually, it's not their fault or so. Because the definition of AI in my opinion, is not very sharp. By the definition that Dries gave, a calculator could even be called an artificial intelligence because it takes over some intelligence of a human being and puts that in a simple device. But what is changing a lot in the recent years is, with the advent of deep learning. And sometimes deep learning is put equal to artificial intelligence. And there, I think it's not true. Deep learning is just a subset of artificial intelligence. But deep learning makes applications possible that were not possible before, getting to a level of human intelligence that is really high actually. And that's something that is changing over the last years, I would say.
Pia Abola:
James, do you have anything you want to add?
James Grayson:
To be clear, I'm definitely a bit more on the application side of artificial intelligence and seeing how it's used with our customers. But I think the thing that really intrigues me is the interpretation of that multivariate data. It's taking in, creating algorithms that can, "Look at something and interpret something as a human would." For UgenTec, thinking of the simple shape of a sigmoidal curve from a PCR amplification and being able to look at that from a feature perspective, instead of just a couple of analytical data points. To be able to look at all the different variables and variants that can go into that and come up with a very human-like decision that could distinguish between something that's either positive or negative, but also intuit, am I seeing some sort of artifact that could be from chemistry or instrument? And that to me is I think very applicable and something that's a strength for UgenTec customers for sure.
Pia Abola:
Can you tell me how is AI helping hospitals understand their data? what is it that's going on?
Dries Hens:
From the LynxCare perspective, what our algorithms do is we mimic a human data extraction process. So today the standard of data extraction within a hospital record, for example, is going through medical records, looking for specific data points and extract them to a structured data format. The problem, however, is that it's a manual process. And if you want to become a data-driven player, and as everybody knows that good structured data becomes really important in the next decade within healthcare in order to transform, for example, to value based healthcare systems. It's not possible today with the manual processing power to do all these tasks. So that's why our AI, so our targeted data-mining models, they look with the same eyes, for example, at data as nurse in cardiology would look to cardiology records. So we train an algorithm to look with the same glasses. So the same glasses that when she opens a medical record, she knows, for example, for a cardiology patient in heart failure, which data points are relevant. And that's why we train algorithms the same way that a human process would do.
Pia Abola:
So can you tell me a little bit more about structured data and what you mean by that?
Dries Hens:
Structured data is in a tabluar format and like Excel, for example, is a structured data. And you have to look at a hospital information system is like everyone's personal computer. So you will have PDFs, you have text files, you have excels within your desktop. And imagine that you have to have a... You have a question where you need to combine all those data sources in order to answer the question. The problem, however, is that they are in different data formats so your word documents are unstructured. So they are just flat texts. You have an Excel with CME or structured data, which is more easily to analyze, of course. But a hospital information system is just the same buildup. So you have different Excels, different PDFs, words, and they all have really relevant information if you look at clinical outcomes. And that's why you need to have an algorithm that can process independent on the format.
Pia Abola:
Oh, yes. So what you're saying is that the unstructured nature of data is a challenge that AI helps address. James, Dirk, do you see that as well?
James Grayson:
UgenTec is using the same sort of methodology where we're using our artificial intelligence to be able to look across different instrument platforms to pull in data from these different formats and interpret it in the same way to create structured outputs and reportable results. So being able to be agnostic across processes or instruments or different inputs that might come from chemistries or extraction processes in our case, and how that affects the outcome or the result, and combining that into a unified, structured, reportable result.
Pia Abola:
But Dirk, I would imagine that the structured information on a medical image is a little bit different. How do you pull information out of that? What do you look at and how do you codify that?
Dirk Smeets:
Yes, that's a good point. And I think we try to mimic somehow what a radiologist would be doing. And when a radiologist looks at an image, he tries to detect certain patterns. On the one hand, patterns that are different from what a normal image should look like, and based on that information, he can help with the diagnosis or she can help with the diagnosis. And this is definitely something that AI can already do for certain patterns. Detection, for example, of bleed inside the brain is something that can already be done by AI algorithms or detection of certain abnormalities in the lungs to help with the detection of COVID-19. What AI can do on top of that, what is more difficult for radiologists, is do exact quantification of what is different from a normal human being. But also what is very difficult with the text that radiologists generate, of course there is, and that's maybe a segue also to Dries, what LynxCare can do out of these natural language reports, extract, again, very valuable information. So here is where two AI applications come together for enriching the data that is extracted from a medical imaging.
Pia Abola:
Dries, did you want to follow up on that?
Dries Hens:
Yeah, I completely agree with Dirk. So you directly see how different the AI companies, you have UgenTec, you have Icometrix, you have LynxCare and we have a little bit the same mission and vision in terms of what we want to do within the healthcare ecosystem, but we have a different core business and a different approach. But what I want to say for COVID-19, huge amounts of data are being generated since the beginning of the episodes. But if we really look, and everybody in the room will probably agree on this, is that the insights generated are really not of the level we thought we would be at the moment.
Dries Hens:
If you look from a governmental perspective in Belgium, but also in Europe, you see that really rich, granular insights are still lagging behind in the healthcare ecosystem. And all these three companies that you have here today, we all are enabling our technology to make sure that in the future, this won't be the case and that we can directly leverage AI applications in order to improve first of all, the insights, but secondly as well, that we can have a direct benefit on patient outcomes and this is what we want to fight for, to be honest, yeah.
Pia Abola:
Dries, I wanted to ask you about, how is your platform really helping with COVID care?
Dries Hens:
Yeah, exactly. So, every patient based on this medical records, you can order them into specific groups based on the comorbidities they have, age, sex, and you have other relevant data points like BMI and so on that are relevant for the outcome of a COVID-19 infection. And being able to have all those data points structured into one common data platform where you have analytics on top, enables physicians now to easily see based on their own patient population, "What is the outcome that I can expect with that patient?" And this is something that is of course of utmost value if you are looking at capacity problems, if you are looking at, for example, we may have vaccine in January, which types of patients should we give the vaccination to, and what outcomes can we expect on the longer term as well?
Pia Abola:
Dirk, I wanted to turn it back to you and ask you a similar question. How is your imaging platform helping healthcare providers give better care?
Dirk Smeets:
Yeah, that actually depends on let's say the phase of the pandemic, as well as the location where we are in the world because the needs are different. If I focus first on the first wave, in many countries there was a lack of capacity to test patients because there were not enough PCR tests, or they could not be executed fast enough. And in this scenario, CT could be used as an alternative for a diagnosis. The CT is looked at and it's quite sensitive to pick up COVID-19. It's not so specific so it means other conditions have very similar appearance on the CT. But when we are in the endemic phase of the pandemic, meaning there is a large occurrence of people having the disease, you can assume that most of the people having symptoms are COVID-19 patients. Now during the second wave or in some countries, the third wave, the lack of capacity is not an issue anymore.
And CT, the role of CT as a diagnostic is not present anymore. Now we look at CT much more from the angle that it can help to understand better patients that need to be hospitalized. We can look inside the lungs of the patient and see how much tissue is affected. And based on that, we have extra information that help us to triage a patient, whether the patient can be sent home for quarantine, whether the patient needs to be admitted to the hospital, or even whether an ICU admission is required. Of course, this information extracted from the CT image should always be complemented with the clinical symptoms of the patient. Actually, these are even more important, but the CT information gives the in insights to be more sensitive and more specific to help the triage further.
Pia Abola:
James, I was hoping you could tell us any of your insights on testing and tell us about how UgenTec is helping support SARS-CoV-2 testing.
James Grayson:
So having the ability to scale up, maintain resources, both being able to support through standardization the ability to bring on new staff that doesn't have to be so qualified because we have the artificial intelligence in place to do the analysis interpretation, and get out results in almost real time from the minute those samples come off the instrument, to being able to get a result, publish that result using automation and intelligence and get that to the point where it can be reported back to the patient. And it's that ability to identify someone with COVID-19 and then take action. And by reducing time to result, I think has been a significant contribution for UgenTec in the pandemic.
Pia Abola:
And to take it back at a more general level, what I'm hearing is artificial intelligence is doing multiple things. It's helping get answers faster, it's really speeding how we can as a community or as medical professionals respond to the pandemic, but it's also helping develop deeper insights and make sense out of information that we really couldn't before, which is great. What are the challenges? What do you see as the challenges in implementing artificial intelligence, either generally or during the pandemic?
Dries Hens:
Yes. Challenges with AI applications and what we've seen as well with our company LynxCare, where we work with different hospitals, especially in the beginning, I think trust of the data is of enormous importance. If you empower your end users and give them the availability of full transparency and always being able to go back to the original source, this is of major importance in terms of trust. And if you're using an AI application within the healthcare ecosystem where your main clients are hospitals and you're handling quite sensitive data, then trust is quite key in order to scale up your company..
Pia Abola:
So I think part of the problem is that a lot of these artificial intelligence algorithms are in fact black boxes. And sometimes they make decisions and the user who is not versed in artificial intelligence and programming really can't figure out, "Why did it make that decision? Should I trust it?" Is that what you're saying?
Dries Hens:
Yes, exactly. I guess data trust comes first of all, by trusting the data or the decision that is being taken by the algorithm. Secondly, as well is if you speak, for example, about end-users and for Icometrix and for LynxCare, the end user often is for example, a physician. It should be also made simple. Simplicity is quite key if you want to trust data. A physician should be able to easily interpret certain result as well as he has to be able to have a simple but easy view on what is being done. And if you make it too complex, then they will never trust your solution because they simply can't work with it. And this is of utmost importance, yeah.
Pia Abola:
Yeah. Dirk, I am sure you must encounter this as well.
Dirk Smeets:
Yeah, absolutely. And I would even add some other big challenge in the development of artificial intelligence. So Dries discussed the challenge of bringing AI into a routine setting, generating the trust with the clinicians. But also in the development, there were some challenges that we have experienced because in, let's say, more complex AI, it's based on training data. Training data that need to be annotated. And here are two big challenges. First is getting the training data and training data in the medical field are also medical data. The second part is get annotations. These annotations need to be smart, need to be intelligent.
We were lucky that we could, in the beginning, have a lot of radiologists helping us by annotating all of these scans. We recruited over 50 radiologists all around the world that helped us to build very quickly that database. And that was luckily possible because also the medical profession understood that when joining forces with technology companies, they could build solutions that could help the medical profession further. So these are definitely also the challenges that I wanted to add that's often not seen, but it's there. And maybe there's also some regional differences. In Europe, you have very specific legislation, in US there's different legislation, in other continents there might be not real strong privacy legislation, and it might be easier to build AI algorithms like these.
Pia Abola:
I think AI has been there a little bit more in diagnostics, but are there still adoption challenges or what are the... Are there other challenges that you face with your end users?
James Grayson:
I primarily agree with the guys here. When you first take it from the trust perspective, how do you build trust around potentially a black box? And we've done a lot of work to allow for that trust to be built. And we have various toggles in our platform that allow you to almost like, "Okay. I have confidence in the step I'm seeing, I'm now willing to throw the toggle and go to the next level of automation, go to the next level of trust." And, so we've as a company, I think, been very aware of how end-users might perceive that experience and whether or not it's something that they're willing to just accept. And, yeah, most end-users aren't willing to just say that, "This is going to be better than a human eye, or this is going to be better than my current process." So we allow them to go through that journey and get to the point where they're willing to personally go into the software, into the configuration settings and throw that switch for full automation if they want to.
Dirk Smeets:
What I could add here is something that is particularly interesting in, let's say, the recent years in the research in the artificial intelligence is what it's called explainable AI. It's quite resource intensive, but nowadays there are some techniques that make that black box less of a black box. And this is something that we, in the last version of the software analyzing CT images, also have implemented. We call them the explainability maps. And actually, these are visual representations of what happens deep inside the network of the artificial intelligence. In our case, for example, we show which lesions, which hotspots in the image that we found that made the algorithm decide it's a COVID case or it's a non COVID case. And if that region, that area, that pattern is corresponding with how a radiologist would see that, he gets confidence.
Dries Hens:
But what we have to discuss as well as is for healthcare institutions, especially hospitals, implementing AI applications or software companies are something new and what we've seen as well in recent year, to be honest, is that a lot of hospitals are setting up a data governance structure. So they can not only handle specific data requests coming from AI companies, but also can maximize the value that AI applications can bring towards their institution.
Pia Abola:
Where do you see AI going in healthcare and the biomedical sciences? What's next? James, do you want to go first?
James Grayson:
Sure. For us, it's truly expansion, right? I think we've all been fixated on COVID. I mean, we're talking about the pandemic, this is the Countering COVID podcast. But eventually this will wind down. So it's, how do we move forward? And for us, it's looking at taking COVID into respiratory panels. How do we support complex analysis and viable reporting when COVID is a member of a multi-target respiratory panel? And how do we... With COVID, we've also seen the expansion of technology with modality for analysis. For us we're seeing where it might've been primarily real-time PCR in the past. Now we're seeing applications of EndPoint PCR. We're seeing isothermal amplification that we need to manage and it's...
James Grayson:
At least for UgenTec, it's how do we take this from COVID to making sure that we get back to our core expertise and being able to do a respiratory panel analysis along with all the other diagnostics? Being able to support the new technologies that have come online and become accepted because of COVID and really grow from there. And then also, I mean, just generally speaking, being ready for the next pandemic, when and if that ever happens. I think everyone here has gained massive experience on how to be ready for some other big event that could come down the line. And I think that's been critical and important for UgenTec.
Dirk Smeets:
Yeah. Actually, on that last one, James, there is some very interesting initiative taken by the OECD, The Organization of Economic Cooperation and Development. They have an initiative that's called The Global Partnership on AI Pandemic Responses, where they try to gather all the learnings of AI from this pandemic and learn from that to be more efficient in the next pandemic. And as Icometrix, we are also contributing to that partnership and we hope that as a society we can learn. We have seen now some limitations of how quickly rolling out AI algorithms. We discussed them earlier in this podcast, like the training, the privacy issues, the annotations, the quality of that. So these are issues that are tackled by the global partnership. And I believe that can make a difference for future pandemics. And it's good that there is this high level global initiative for this.
Dries Hens:
Yeah. And that's also a good point from our end. So we initiated, for example, the COVID-19 Project, we initiated it in Belgium, but we are now part of the EDEN Project, which is... EDEN is The European Health Evidence and Data Network. And together with 28 data partners in Europe, we are collecting the same data in the same format. So you can directly see on a European scale as well that different companies are really collaborating into increasing the insights, as well as just highlighting the lessons learnt and those hopefully will really result in specific guidelines for the future. And this is where I really hope that our companies as well... It's beautiful what we did with Icometrix and UgenTec. We really demonstrated that that AI is here to stay and can really enable value towards healthcare as an enabler for physicians to manage specific questions. And I guess being part of a more European network and pull it more on a global scale will only make sure that in the future for our next pandemic, everyone will be ready, hopefully, yeah.
Pia Abola:
I think that was a wonderful way to end. We hope you enjoyed today's episode on Countering COVID with AI. In our next episode, we'll speak with scientists from Combinati, a company that's using their digital PCR platform to support SARS-CoV-2 testing of wastewater for community surveillance of COVID. Thank you.